TerraMosaic Daily Digest: Feb 3, 2026
Daily Summary
This digest synthesizes 298 selected papers and focuses on landslide process mechanics and slope evolution, flood generation, routing, and hydroclimatic forcing, coastal and submarine hydro-geomechanics. Top-ranked studies examine satellite and LiDAR-based deformation monitoring, coastal and submarine hazard coupling, and landslide susceptibility mapping.
Across the full set, evidence converges on mechanism-constrained analysis with operational relevance, especially for infrastructure-focused hazard performance and risk, fragility, and resilience quantification. The strongest contributions pair interpretable process evidence with monitoring or forecasting workflows that support warning design and risk prioritization.
Key Trends
- Landslide studies increasingly resolve process chains: Contributions connect triggering conditions, slope deformation, and mobility outcomes, improving the basis for warning thresholds and scenario testing.
- Flood analyses are becoming event-specific and process-based: Papers emphasize precipitation structure, antecedent wetness, and catchment controls rather than static hazard descriptors.
- Coastal and submarine hazards are treated as coupled systems: Wave, mass-transport, and shoreline processes are analyzed together with engineering implications.
- Infrastructure-facing outputs are increasingly decision-ready: Asset performance is evaluated with uncertainty-aware frameworks to support mitigation and maintenance prioritization.
- Risk studies move beyond hazard mapping to consequence pathways: Vulnerability, fragility, exposure, and recovery metrics are integrated to compare interventions under compound hazards.
Selected Papers
This digest features 298 selected papers from 1406 RSS items analyzed across multiple journals. Each paper has been evaluated for its relevance to landslide and broader geohazard research and includes links to the original publications.
1. Advancing Near‐Real‐Time Flood Inundation Mapping in Australia
Core Problem: Minimizing loss of life and socio-economic impacts from floods in Australia by providing near-real-time (NRT) flood extent and depth maps, which requires leveraging advanced computing and diverse data sources.
Key Innovation: Developed a flood monitoring workflow providing NRT 5-m spatial resolution flood extent and depth maps using airborne LiDAR, gauge data, coupled hydrological/hydrodynamics models, and satellite observations, demonstrating its effectiveness and transferability for strengthening community resilience.
2. Tsunamigenic potential of unstable masses in the Gulf of Pozzuoli, Campi Flegrei, Italy
Core Problem: The potential hazard posed by gravitational instabilities (landslides) and their interaction with water to trigger tsunamis in the Campi Flegrei volcanic area has received little consideration, despite its dense coastline.
Key Innovation: Reconstructed and simulated a set of four landslide-tsunami scenarios (one subaerial, three submarine) through a sequence of numerical codes, accounting for all phases of the tsunami process, providing insights into tsunami energy distribution, identifying most affected coastal stretches, and exploring the influence of dispersion and resonance effects.
3. From typhoon rainfall to slope failure: optimizing susceptibility models and dynamic thresholds for landslide warnings in Zixing City, China
Core Problem: Existing landslide warning systems inadequately capture the distinct rainfall dynamics of typhoon-specific rainfall-induced landslides, which pose critical hazards in mountainous regions.
Key Innovation: Proposed an integrated framework combining optimized susceptibility predictions (using buffer-based negative sampling and variable weighting with SVM, achieving F1-score: 0.859, AUC: 0.914) with dynamic rainfall thresholds tailored to typhoon patterns (identifying 24h intensity + 7d antecedent rainfall as optimal trigger), demonstrating high spatial efficiency for landslide warnings.
4. Bedrock ledges, colluvial wedges, and ridgetop wetlands: characterizing geomorphic and atmospheric controls on the 2023 Wrangell landslide to inform landslide assessment in Southeast Alaska, USA
Core Problem: Several fatal landslides in Southeast Alaska highlight the need to advance understanding of regional geomorphic and atmospheric controls on triggering events and runout behaviour, particularly for large, long-runout events.
Key Innovation: Characterized the geomorphic, hydrologic, and atmospheric conditions contributing to the 2023 Wrangell landslide using field observations, sequential airborne lidar, geotechnical analyses, and climate data, suggesting that the sequencing of rain- and snow-dominated storms, geologic controls on colluvium production/accumulation, and ridgetop hydrology contributed to initiation and runout, informing regional hazard assessment.
5. Review article: Deep learning for potential landslide identification: data, models, applications, challenges, and opportunities
Core Problem: The increasing frequency and severity of landslide hazards due to global climate change and human activities necessitate improved early identification, but deep learning applications face challenges like data imbalance and limited model generalization.
Key Innovation: Systematically reviewed over 400 studies on deep learning for potential landslide identification, summarizing data sources (satellite, airborne, ground-based), model types (image, time series analysis), applications (rainfall-, earthquake-, human activity-, multi-factor-induced landslides), challenges, and future directions, suggesting integration of knowledge-driven and data-driven approaches.
6. Dynamic analysis of flowlike landslides at Brienz/Brinzauls, Graubünden, Switzerland
Core Problem: Accurate forecasting of the risk posed by catastrophic failure of rock slopes requires estimates of potential impact area and emplacement velocity, and a wider range of landslide classes beyond rock avalanches need to be considered.
Key Innovation: Derived and implemented a GPU-accelerated numerical model capable of simulating emplacement velocities on the order of meters per day, performing forensic back-analysis of two flowlike rock slope failures at Brienz/Brinzauls, highlighting the critical role of path material and moderate changes in source material lithology in controlling emplacement behavior.
7. Autonomous and efficient large-scale snow avalanche monitoring with an Unmanned Aerial System (UAS)
Core Problem: Current and accurate information about the location and extent of released avalanches is critical for public safety but difficult and expensive to obtain in remote locations, and autonomous UAS flight in mountainous terrain remains challenging.
Key Innovation: Presented a proof-of-concept system capable of safely navigating and mapping avalanches using a fixed-wing Unmanned Aerial System (UAS), demonstrating effective and safe navigation in steep mountain environments while maximizing map quality and efficiency, with potential to significantly impact avalanche warning, mitigation planning, and hazard mapping.
8. Predicting the amplitude and runup of the water waves induced by rotational cliff collapse, considering fragmentation
Core Problem: Cliff collapses in small lakes and reservoirs induce powerful waves threatening offshore infrastructure, but previous studies focused on granular slides, and understanding waves induced by rotational cliff collapse with fragmentation is needed.
Key Innovation: Experimentally and numerically investigated waves induced by rotational cliff collapse (considering fragmentation), finding that it produced 28%–42% higher wave amplitudes than equivalent granular slides due to acute impact, and developed machine learning models to predict wave amplitude and runup with high R2 values, highlighting the cumulative 90% contribution of impact velocity, cliff height, and number of fragments to wave amplitude.
9. Mechanisms and scenarios of the unprecedent flooding event in South Brazil 2024
Core Problem: The unprecedented May 2024 flooding event in South Brazil, particularly impacting complex river–estuary–lagoon systems, requires a detailed hydrodynamic assessment to understand its governing mechanisms and evaluate potential hydraulic flood control interventions for future preparedness.
Key Innovation: First detailed hydrodynamic assessment of the May 2024 South Brazil flood using a validated model (with SWOT satellite observations), investigating the main mechanisms (e.g., Taquari and Jacuí river contributions, synchronization of peaks) and assessing the limited effectiveness of proposed hydraulic flood control interventions for the Metropolitan region of Porto Alegre.
10. Integrating geospatial data and hybrid machine learning models for landslide susceptibility assessment in semi-arid environments
Core Problem: Landslides are destructive natural hazards with severe impacts, and accurate mapping and prediction of landslide-prone areas are essential for effective management, especially in semi-arid environments.
Key Innovation: Proposes integrating geospatial data and hybrid machine learning models for landslide susceptibility assessment in semi-arid environments.
11. An abandoned road as a debris trap: Estimating debris-supply rate from steep slopes based on UAV–LiDAR DEMs
Core Problem: Decadal-scale observation of debris supply, an important factor controlling the frequency of debris flows in steep headwater streams, has been technically difficult.
Key Innovation: Proposed a new method using abandoned roads as debris traps and UAV–LiDAR DEMs to estimate debris-supply rate by rockfall into debris-flow prone channels over a decadal timescale, estimating 70–93 m3/yr for a given source area and indicating runoff-generated debris flows may occur every few decades.
12. Robust sequential pixel offset tracking (RS-POT): A novel monitoring approach for landslides with long-term large deformations
Core Problem: Existing time-series SAR pixel offset tracking (POT) approaches struggle with precise long-term monitoring of landslides experiencing large deformations, often underestimating deformation in high-gradient regions.
Key Innovation: A Robust Sequential Pixel Offset Tracking (RS-POT) method that derives short-term deformations using single-reference POT and sequentially integrates them through a robust fusion strategy, ensuring long-term continuity and high accuracy for large landslide deformations, outperforming PO-SBAS.
13. Influence of interface flow field on slope deterioration mechanism under freeze–thaw cycles: insights from physical model testing
Core Problem: Existing studies on freeze-thaw-induced slope deterioration often neglect the effects of forced convection heat transfer at the atmosphere-soil interface, which significantly modifies thermal boundary conditions and influences internal thermo-hydro-mechanical (THM) responses.
Key Innovation: Developed a novel freeze-thaw cycling apparatus to simulate convective heat transfer boundaries and conducted physical model tests, revealing that forced convection accelerates freezing front migration and crack initiation, and that the direction of convection controls the slope failure process, leading to distinct sliding patterns.
14. Deformation mechanism of landslide caused by wave scouring considering water level rise
Core Problem: Understanding the evolutionary process and instability mechanism of landslide deformation caused by wave scouring, especially under the influence of water level rise, in reservoir areas like the Three Gorges, is critical.
Key Innovation: Revealed that large landslide deformation is caused by the combined effect of reservoir water rise and wave scouring, with water level increase significantly enhancing erosion and triggering instability, and proposed a four-stage deformation process for such landslides.
15. The Effects of Planting Structure on Groundwater Depletion and Optimization Strategies in the North China Plain
Core Problem: Quantifying the effects of planting structure changes on groundwater depletion and developing optimization strategies to alleviate severe groundwater depletion, especially in high groundwater stress zones (HGSZ), which are often overlooked.
Key Innovation: Developed a groundwater stress index (GWSI) and a GWSI-based optimization model, revealing that current planting structures lead to unsustainable groundwater use, and demonstrating that adjusting planting structures (e.g., converting to single-season cropping, shifting areas) effectively alleviates depletion.
16. Variational and Monte Carlo Methods for Bayesian Inversion of Dynamic Subsurface Flow Simulations Using Seismic and Fluid Pressure Data
Core Problem: Accurately predicting the future performance of subsurface fluid reservoirs and estimating their properties (e.g., permeability) from noisy seismic and fluid pressure measurements involves solving challenging inverse problems, requiring robust methods for posterior probability distribution estimation and uncertainty quantification.
Key Innovation: Compares Monte Carlo and variational inference methods (ADVI, SVGD, sSVGD, PSVI) for Bayesian inversion of dynamic subsurface flow simulations. It demonstrates that PSVI (physically structured variational inference) achieves a good balance of accuracy and computational efficiency for estimating reservoir permeability, particularly for CO2 storage applications, outperforming other methods in specific aspects.
17. Scaling laws for rockfall impact fragmentation emerging from diverse lithologies
Core Problem: Predicting debris evolution in rockfall events is a persistent challenge due to the stochastic nature of impact-induced fragmentation.
Key Innovation: Introduces a discrete element framework revealing a universal Weibull scaling law for fragment size distributions in rockfall, independent of lithology or initial kinetic energy. This provides a robust predictive link between impact mechanics and structural resilience, informing hazard mitigation strategies.
18. Thalia: A Global, Multi-Modal Dataset for Volcanic Activity Monitoring
Core Problem: Progress in automating and enhancing InSAR interpretation for volcanic activity monitoring using deep learning has been limited by the scarcity of well-curated, high-resolution, multi-source, and multi-temporal datasets for geohazard assessment.
Key Innovation: Introduces Thalia, a global, multi-modal dataset for volcanic activity monitoring, enriching existing data with higher-resolution InSAR products, topographic data, and atmospheric variables, accompanied by expert annotations and a comprehensive benchmark for classification and segmentation.
19. A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics
Core Problem: High-fidelity numerical models for coastal and riverine shallow water dynamics are too computationally expensive for real-time forecasting, while lower-cost approaches struggle to generalize to out-of-distribution conditions, hindering infrastructure planning and climate adaptation.
Key Innovation: MITONet, a novel autoregressive neural emulator, employs latent-space operator learning to efficiently approximate high-dimensional numerical solvers for 2D shallow-water equations. It displays consistently high predictive skill for tide-driven and riverine flow dynamics, with significant computational speedups and generalization to unseen parameters.
20. Seismic site amplification assessment in the Dakwäkäda (Haines Junction) area of Yukon, Canada, from probabilistic inference of passive seismic measurements
Core Problem: Lack of knowledge about local site properties in a tectonically active region (Dakwäkäda, Yukon) hinders accurate assessment of earthquake shaking intensity and duration (site effects), which are crucial for natural hazard mitigation.
Key Innovation: Probabilistic inference of 1D subsurface shear-wave velocity from passive seismic measurements at 14 sites, including rigorous uncertainty quantification, to classify sites and estimate linear site amplification factors, revealing spatial variability linked to hydrologic/cryospheric processes and highlighting the mitigating role of permafrost and the need to study thaw effects.
21. Tsunami-Like Solitary Wave Overtopping on Vegetated-Foreshore Sea Dikes: Characterization and Prediction
Core Problem: Existing empirical formulas for tsunami-like wave overtopping on vegetated foreshores are limited and do not account for coupled effects, hindering effective coastal protection design.
Key Innovation: Developed a 2D numerical model and robust predictive models (MNLR, ANN) for wave overtopping discharge, demonstrating that vegetation significantly reduces overtopping and steeper foreshores enhance attenuation, providing guidance for nature-based coastal protection.
22. Multiscale modeling for coastal cities: addressing climate change impacts on flood events at urban-scale
Core Problem: Accurate flood hazard assessment in coastal cities requires high-resolution simulations to capture the influence of local topography and infrastructure, especially where global DEMs are inadequate, and to bridge scales from regional climate projections to urban flood impacts.
Key Innovation: Developed an integrated modeling framework that employs a novel non-standard downscaling approach to translate large-scale atmospheric outputs from EURO-CORDEX regional models into high-resolution (2–20 m) urban flood simulations of storm surges, wave climate, and river discharge, enabling detailed analysis of changes in flooded areas and volumes under future climate scenarios.
23. Large discrepancies between event- and response-based compound flood hazard estimates
Core Problem: Most flood hazard assessments (event-based) neglect information about the temporal and spatial variability of flood drivers and processes, leading to large discrepancies compared to response-based approaches that account for these factors.
Key Innovation: Compared event- and response-based approaches for compound flood hazards in Gloucester City (NJ), finding that compound events with return periods less than 20 years can produce 100-year flood depths due to specific temporal and spatial characteristics (e.g., prolonged high coastal water levels, higher rainfall over urban areas), emphasizing the need to incorporate these variabilities for robust estimates.
24. Land subsidence dynamics and their interplay with spatial and temporal land-use transitions in the Douala coastland, Cameroon
Core Problem: Knowledge of the drivers and impact of coastal subsidence in the Douala coastland (DCL) remains limited, despite the region experiencing alarming rates of coastal erosion, frequent flooding, and significant land loss.
Key Innovation: Used Sentinel-1 C-band InSAR datasets (2018–2023) to quantify vertical land motion (VLM) in the DCL, revealing subsidence rates ranging from −17.6 to 3.8 mm yr−1 (mean 2.7 mm yr−1), and demonstrated an inverse relationship between subsidence rates and the timing of land-use and land-cover (LULC) changes into urban areas, highlighting the impact of urban expansion on present-day subsidence.
25. Quantifying the influence of coastal flood hazards on building habitability following Hurricane Irma
Core Problem: Estimating which buildings will become uninhabitable due to coastal flood events like tropical cyclones is critical for creating resilient communities, but current estimations need increased accuracy.
Key Innovation: Developed habitability functions to quantify the relationship between hydrodynamic hazards (e.g., maximum flood depths modeled using Delft3D-FM and SWAN for Hurricane Irma) and the probability of a building becoming uninhabitable, finding maximum unit discharge to be the best predictor and noting that wooden structure habitability is significantly influenced by hazard level, unlike concrete structures.
26. Assessing the spatial correlation of potential compound flooding in the United States
Core Problem: While understanding of compound flooding has improved, no studies to date have assessed the spatial correlation of compound flooding, which is crucial for comprehensive risk assessments.
Key Innovation: Developed a framework to capture dependence between coastal total water level and river discharge across US coastlines, stochastically modeling 10,000 years of spatially-joint extreme events, revealing high spatial correlation of potential compound flooding on the US West coast (around 50% of compound events affecting multiple locations simultaneously) and weaker correlations on the East and Gulf coasts, highlighting the importance of accounting for spatial dependence.
27. Evaluating the feasibility of scaling the FIER framework for large-scale flood inundation prediction
Core Problem: Traditional flood forecasting methods face computational and data challenges for large geographic areas, especially in data-scarce regions.
Key Innovation: Developed a novel approach to scale the data-driven FIER framework for large-scale flood inundation prediction by using watershed boundaries to create and mosaic individual models, demonstrating improved accuracy for predicting flood extents.
28. Generating Boundary Conditions for Compound Flood Modeling in a Probabilistic Framework
Core Problem: Generating a large enough set of storm events for boundary conditions remains a challenge for compound flood models, limiting comprehensive probabilistic risk assessments.
Key Innovation: Introduced a statistical framework to generate many synthetic but physically plausible compound events (storm-tide hydrographs and rainfall fields) as boundary conditions for dynamic compound flood models, demonstrating its effectiveness and highlighting the importance of accounting for mean sea level and tidal variability.
29. Probabilistic hierarchical interpolation and interpretable neural network configurations for flood prediction
Core Problem: Improving the accuracy and providing probabilistic understanding for neural network models used in flood prediction.
Key Innovation: Developed and benchmarked two interpretable neural network models (N-HiTS and N-BEATS) with a probabilistic multi-quantile objective for flood prediction, demonstrating improved accuracy and providing probabilistic predictions for flooding events.
30. Refined collapse susceptibility assessment in Tonghua city based on collapse source area data and multi-model coupling
Core Problem: Tonghua City, with its steep terrain, heavy rainfall, and intensive human activity, is highly prone to collapses that endanger settlements and infrastructure, necessitating a refined susceptibility assessment.
Key Innovation: Proposes a refined collapse susceptibility assessment method for Tonghua City based on collapse source area data and multi-model coupling.
31. Fuzzy-decision trees models for flood hazard modeling in the Danube Delta
Core Problem: Assessing flood susceptibility in deltaic areas requires innovative methodological frameworks.
Key Innovation: Proposes an innovative methodological framework for assessing flood susceptibility in deltaic areas by integrating fuzzy logic with three decision tree algorithms: CART, Random Forest, and Gradient Boosting.
32. Probabilistic models in rock slope kinematic analysis employing the reliability engineering approaches and considering the variability of rock joint orientations
Core Problem: The variability of rock joint orientations significantly influences slope failure type, making probabilistic methods crucial for kinematic analysis and assessing slope stability under uncertainty.
Key Innovation: Integrated probability kinematic analysis with reliability engineering methodologies (RBD, ETA, FTA) to assess rock slope stability under joint orientation uncertainty, examining the effect of friction and lateral limit angles using RSM, and demonstrating consistent output and effective failure probability estimation.
33. Seismic response and failure mechanism of pile foundations at different relative positions and rock-socketed depths on deep deposit slopes
Core Problem: Understanding the seismic response and failure mechanisms of rock-socketed pile foundations at different relative positions and rock-socketed depths on deep deposit slopes (DDPs) in southwestern China, where numerous deposit landslides and pile failures occur following earthquakes.
Key Innovation: Conducted shaking table tests to assess the impact of seismic action on pile foundations, revealing that deposit nonlinearity, pile position, and rock-socketed depth significantly influence seismic response, deformation, and forces, leading to a failure mode characterized by shallow sliding and providing recommendations for seismic design.
34. Seismic fragility of CAP1400 nuclear power plant consisting of steel-concrete shield building under mainshock-aftershock sequences
Core Problem: Lack of comprehensive seismic fragility analysis for CAP1400 nuclear power plants (NPPs) under mainshock-aftershock (MS-AS) sequences, which can cause more severe damage than single ground motions.
Key Innovation: A comprehensive seismic fragility analysis of CAP1400 NPPs under MS-AS sequences, demonstrating increased damage from aftershocks and the significantly lower fragility and superior seismic performance of steel-concrete (SC) shield buildings compared to conventional reinforced concrete (RC) shield buildings.
35. Full-Process simulation and risk assessment framework for compound flooding in inland cities with high external inflow
Core Problem: Inland cities with high external inflow (HEI) face increased flood risks from compound urban flooding (combined external inflow and local surface runoff), but current research lacks emphasis on these compound mechanisms and risk progression.
Key Innovation: A multi-module dynamic coupling framework for comprehensive modeling and risk assessment of compound urban flooding, integrating hydrological, 2D surface hydrodynamic, 1D drainage network, and 1D river hydrodynamic modules, along with a dual-indicator risk assessment, revealing how external inflow and elevated river stages amplify different flood risks.
36. Effects of the joint aperture and persistence on the shear behavior of coplanar non-persistent jointed rock masses and an improved Jennings shear strength criterion
Core Problem: The shear behavior and failure mechanisms of non-persistent joints, critical for jointed rock mass stability, are complex and jointly governed by geometric parameters like joint aperture and persistence, requiring a more comprehensive shear strength criterion.
Key Innovation: Performed direct shear tests and FEM-CZM simulations to investigate the combined effects of joint aperture and persistence on rock mass shear behavior, deriving an improved Jennings shear strength criterion that incorporates the weakening effect of joint aperture.
37. Study on longitudinal seismic fragility and resilience of shield tunnels crossing soil–rock interfaces
Core Problem: Understanding and assessing the longitudinal seismic fragility and resilience of shield tunnels, especially those crossing complex soil-rock interfaces, is crucial for managing seismic risk in infrastructure.
Key Innovation: Analyzed the seismic fragility and resilience of shield tunnels crossing soil-rock interfaces using the generalized response displacement method and incremental dynamic analysis, revealing that interfaces amplify seismic effects and near-fault pulse-type motions reduce performance, and quantifying the influence of key factors on economic loss and resilience.
38. Damage mechanism and optimization design of wall-type portal tunnels underseismic action: Seismic damage investigation and a theoretical model
Core Problem: The deformation incompatibility between wall-type portals and tunnel linings during seismic events is a significant cause of damage, but theoretical analysis methods for assessing this seismic damage are lacking.
Key Innovation: Developed a theoretical model simplifying the wall-type portal as a concentrated mass and the lining as a Timoshenko beam to evaluate the longitudinal dynamic response of wall-type portal tunnels under seismic action, providing an analytical solution and strategies for optimizing seismic performance.
39. Centrifuge modelling with an equivalent mixed model for liquefaction response of deep sand ground
Core Problem: Physical modeling of liquefaction response in deep sand ground is limited by scaling issues, making it challenging to accurately replicate field conditions in centrifuge tests.
Key Innovation: Developed an equivalent mixed model for centrifuge testing, consisting of dry lead shot, a functional inter-layer, and liquefiable sand, with corresponding design criteria, which effectively replicates small-strain shear modulus, acceleration, settlement, and excess pore water pressure generation during liquefaction.
40. Large Streamflow Differences Between Forested and Urbanized Watersheds in the Energy‐Limited Eastern United States: The Role of Evapotranspiration and Impervious Surfaces
Core Problem: Understanding the hydrological impacts of urbanization, specifically the differences in streamflow and flood generation between forested and urbanized watersheds, and how Nature-based Solutions (NbS) can mitigate urban flooding.
Key Innovation: Quantified large streamflow differences, showing urban watersheds have higher water yield and peak flows (up to 100 times higher during hurricanes) compared to forested ones, and developed conceptual models to explain these differences, emphasizing maximizing evapotranspiration and minimizing impervious surfaces for flood mitigation.
41. Observation‐Driven Forecast of Global Terrestrial Water Storage and Evaluation for 2010–2024
Core Problem: The latency of GRACE/GRACE-FO terrestrial water storage change (TWSC) products limits their utility for real-time and operational forecasting applications, hindering drought and flood risk management.
Key Innovation: Developed a machine learning-based method to forecast GRACE-like TWSC up to 12 months ahead using observational/reanalysis inputs, offering improved accuracy and robustness over existing seasonal forecasts, and providing a real-time TWSC forecast for drought early warning and flood risk management.
42. Refining Decision Boundaries In Anomaly Detection Using Similarity Search Within the Feature Space
Core Problem: Detecting rare and diverse anomalies in highly imbalanced datasets (e.g., APTs in cybersecurity) remains a fundamental challenge, and conventional active learning often fails to exploit feature space geometry for model refinement.
Key Innovation: Introduces SDA2E, a Sparse Dual Adversarial Attention-based AutoEncoder for learning compact latent representations, and a similarity-guided active learning framework with novel strategies (normal-like expansion, anomaly-like prioritization, hybrid) and a new similarity measure (SIM_NM1), achieving superior ranking performance and reducing labeled data by up to 80%.
43. Joint Background-Anomaly-Noise Decomposition for Robust Hyperspectral Anomaly Detection via Constrained Convex Optimization
Core Problem: Most existing hyperspectral anomaly detection methods do not explicitly consider or robustly handle various types of noise (e.g., sparse, stripe noise), which significantly degrades detection performance in real-world degraded HS images.
Key Innovation: A novel robust HS anomaly detection method that formulates a constrained convex optimization problem to jointly decompose background, anomaly, and three types of mixed noise from HS images, achieving comparable accuracy to SOTA on original images and significantly higher robustness under mixed noise.
44. ResSR: A Computationally Efficient Residual Approach to Super-Resolving Multispectral Images
Core Problem: Multispectral imaging (MSI) sensors have wavelength-dependent resolution, limiting downstream analysis, and existing MSI super-resolution (MSI-SR) methods are computationally expensive, hindering large-scale or time-critical applications.
Key Innovation: Introduces ResSR, a computationally efficient, model-based MSI-SR method that decouples spectral and spatial processing into separate branches combined with a residual correction step, achieving comparable or improved reconstruction quality 2x to 10x faster than existing methods without supervised training.
45. Rethinking Efficient Mixture-of-Experts for Remote Sensing Modality-Missing Classification
Core Problem: Multimodal remote sensing classification suffers from significant performance degradation when modalities are missing due to sensor failures or environmental interference, limiting its robustness and generalization.
Key Innovation: Proposes Missing-aware Mixture-of-LoRAs (MaMOL), a parameter-efficient MoE framework that unifies multiple modality-missing cases. It uses a dual-routing mechanism to decouple modality-invariant and modality-aware experts, enabling automatic expert activation conditioned on available modalities, significantly improving robustness and generalization in remote sensing.
46. Dynamic behaviour of biochar-amended soil under cyclic undrained conditions
Core Problem: Railway infrastructure earthwork on weak or problematic soils is challenging under traffic-induced cyclic loads, and conventional soil improvement methods (cement, lime) raise environmental concerns.
Key Innovation: Investigates biochar as a sustainable soil reinforcement, showing that adding 5% biochar notably improved soil cohesion and internal friction angle, enhancing ground stability and accelerating pore water dissipation under cyclic undrained conditions.
47. Flood risks to the financial stability of residential mortgage borrowers: an integrated modeling approach
Core Problem: Little is known about the prevalence and drivers of credit constraints among flood-exposed property owners, despite prior research linking floods to higher rates of financial distress, making recovery uncertain for uninsured households.
Key Innovation: Used a simulation-based approach to estimate the impact of uninsured damage on residential mortgage borrowers' financial conditions over a series of floods in North Carolina, projecting the number of borrowers experiencing credit constraints due to negative equity or liquidity issues, finding 66% of damage uninsured and 32% of affected borrowers lacking sufficient financing for repairs.
48. A Global Ensemble Forecast System (GEFS)-based synthetic event set of U.S. tornado outbreaks
Core Problem: Tornado outbreak risk estimates from observations are limited by meteorological conditions that have occurred in the historical period, necessitating synthetic event sets to represent the full range of possible outcomes.
Key Innovation: Constructed and evaluated a synthetic event set of U.S. tornado outbreaks using Global Ensemble Forecast System (GEFS) environments and a tornado outbreak index, generating over 200,000 synthetic events to estimate 1-in-100-year and 1-in-1000-year outbreak magnitudes (150–250 and 275–400 F/EF1+ tornadoes per day, respectively) and showing robust shifts related to ENSO and trends.
49. Coastal-Cosmo-Model (CCMv1): a cosmogenic nuclide model for rocky coastlines
Core Problem: Understanding the long-term evolution of rocky coasts and reconstructing their histories requires models that can account for complex interactions between exposure, erosion, and sea level, constrained by empirical observations, which existing tools may not fully address.
Key Innovation: Introduction of Coastal-Cosmo-Model version 1 (CCMv1), a modular forward modeling framework that reconstructs coastal histories from in situ cosmogenic nuclide concentrations, integrating community-standard production rate calculations and allowing flexible inversion of platform histories for both eroding and non-eroding coastlines.
50. Trends and prospects of engineering disasters in cold regions: insights from a bibliometric and review analysis
Core Problem: Understanding the trends and future prospects of engineering disasters in cold regions is crucial for sustainable development.
Key Innovation: A bibliometric analysis of 7593 cold region engineering disaster papers (2000–2025) revealed a paradigm shift toward interdisciplinary research integrating permafrost ecology and low-carbon infrastructure.
51. Context matters: practitioners’ perspectives on prepositioning strategies in domestic emergency management
Core Problem: Understanding practitioners' perspectives on prepositioning strategies in domestic emergency management is crucial for effective disaster preparedness and response.
Key Innovation: Focuses on practitioners' perspectives on prepositioning strategies in domestic emergency management, implying an analysis of how these strategies are framed and enacted to improve disaster preparedness and response.
52. Monitoring growth of the wildland-urban interface in 2000 and 2020 in Mediterranean ecosystems with Landsat satellite imagery
Core Problem: Mapping long-term growth of the Wildland-Urban Interface (WUI) to manage wildfire risk is challenging due to the lack of consistent historical data on building locations or density for most WUI mapping methods.
Key Innovation: A novel method using spectral-temporal metrics and neural network regression on Landsat satellite data to consistently map WUI and its growth from 2000 to 2020, accurately distinguishing WUI types by vegetation composition and providing crucial information for wildfire risk mitigation.
53. Quantitative assessment of the reservoir-induced impact on multivariate flood risk via the nonstationary Vine Copula model
Core Problem: Capturing the nonstationary characteristics of flood events in the Yellow River Basin, which are increasingly influenced by both climate change and intensified human activities, and quantifying their impacts.
Key Innovation: A dynamic Vine Copula (DVC) framework integrating nonstationary univariate distributions with time-varying copula parameters to detect and attribute human activities (reservoir operations, urbanization) and climate change impacts on multivariate flood occurrence likelihood, showing reservoir operations have the strongest association with flood risk alterations.
54. Ground failure mechanism for deep tunnel in sandy cobble strata based on the cohesive zone element
Core Problem: Clarifying the ground failure mechanism during shield tunneling in heterogeneous and discrete sandy cobble strata is challenging but crucial for theoretical understanding and practical application.
Key Innovation: Established a numerical analysis method using cohesive zone elements and a dynamic explicit algorithm to simulate tunnel excavation, effectively capturing the progressive failure process and instability range of ground in sandy cobble strata, and clarifying the evolution of collapse and soil pressure arching.
55. Improved Penzien theoretical model for pile groups and dynamic response analysis of nuclear power plants on non-bedrock foundations
Core Problem: The traditional Penzien model for pile groups fails to accurately capture inter-pile interaction, leading to inaccuracies in simulating seismic responses of pile-soil-structure systems, particularly for critical infrastructure like nuclear power plants on non-bedrock foundations.
Key Innovation: Proposed an improved Penzien theoretical model that incorporates the inter-pile soil system to enhance simulation accuracy of seismic responses in pile-soil-structure systems, validated against experimental data, and applied it to analyze nuclear power plant dynamic responses under varying seismic inputs.
56. Effects of the long-term degradation of fiber reinforced cementitious matrix (FRCM) systems on the seismic retrofitting of historical masonry structures
Core Problem: The long-term durability and performance degradation of Fiber Reinforced Cementitious Matrix (FRCM) systems under environmental conditions (temperature, aging) pose a critical challenge for their effective application in seismic retrofitting of historical masonry structures.
Key Innovation: Conducted an extensive numerical investigation to assess the influence of long-term degradation of B-FRCM and G-FRCM systems on the seismic performance of retrofitted unreinforced masonry walls, considering various environmental conditions and structural configurations, providing insights into their overall effectiveness against in-plane failure mechanisms.
57. Stress and Rock Failure Near Salt Bodies: Insights From Field Observations, Kinematic Modeling, and Mechanical Analysis Near Arches National Park, Paradox Basin, Utah
Core Problem: Salt's ductile deformation creates significant mechanical contrasts that perturb the stress field in surrounding rocks, leading to spatially variable local stress states and rock failure, which is critical for subsurface operations and predicting rock stability.
Key Innovation: Integrates field observations, UAV fracture data, and computationally efficient elastic dislocation modeling to simulate and predict spatially variable local stress states and rock failure patterns adjacent to salt bodies, offering a valuable tool for refining subsurface interpretations.
58. CTRIP‐HyDAS: A Global‐Scale Data Assimilation Framework for SWOT‐Derived Discharge Using Synthetic Observations at High Resolution (1/12°)
Core Problem: Improving global discharge predictions, especially in data-scarce regions, by effectively integrating satellite-based observations into hydrological models.
Key Innovation: Developed CTRIP-HyDAS, a global-scale data assimilation framework merging SWOT-derived discharge observations with a high-resolution river routing model, demonstrating widespread improvements in discharge predictions and robust performance even with high observation errors.
59. IceBench-S2S: A Benchmark of Deep Learning for Challenging Subseasonal-to-Seasonal Daily Arctic Sea Ice Forecasting in Deep Latent Space
Core Problem: Current deep learning models for Arctic sea ice forecasting are limited to subseasonal scales, hindering real-world applications requiring daily subseasonal-to-seasonal (S2S) forecasts.
Key Innovation: Introduces IceBench-S2S, the first comprehensive benchmark for evaluating deep learning approaches to extend daily Arctic sea ice concentration forecasting to 180-day S2S scales using a deep latent space framework.
60. VOILA: Value-of-Information Guided Fidelity Selection for Cost-Aware Multimodal Question Answering
Core Problem: Multimodal vision-language systems incur significant costs from retrieving and processing high-fidelity visual inputs, yet most operate at fixed fidelity levels, leading to inefficient resource utilization.
Key Innovation: Introduces VOILA, a Value-Of-Information-driven framework for adaptive fidelity selection in VQA. It uses a two-stage pipeline (gradient-boosted regressor and isotonic calibrator) to estimate correctness likelihood at different fidelities, selecting the minimum-cost fidelity that maximizes expected utility, achieving 50-60% cost reductions while retaining high accuracy across diverse datasets including FloodNet.
61. Experience-Driven Multi-Agent Systems Are Training-free Context-aware Earth Observers
Core Problem: LLM agents struggle in specialized, tool-intensive Earth Observation (EO) domains due to a lack of fine-grained, tool-level expertise, making them unable to reliably configure tool parameters or recover from mid-execution failures.
Key Innovation: GeoEvolver, a self-evolving multi-agent system that enables LLM agents to acquire EO expertise through structured interaction without parameter updates, by decomposing queries, exploring tool-parameter configurations, and distilling successful patterns and failure attributions into an evolving memory bank.
62. Mapping the Unseen: Unified Promptable Panoptic Mapping with Dynamic Labeling using Foundation Models
Core Problem: Open-vocabulary models in panoptic mapping repeatedly produce closely related labels that split panoptic entities and degrade volumetric consistency, hindering robust open-world scene understanding for robots.
Key Innovation: UPPM (Unified Promptable Panoptic Mapping), which leverages foundation models to introduce a panoptic Dynamic Descriptor that reconciles open-vocabulary labels with unified category structure and geometric size priors, resulting in a persistent and promptable panoptic map without additional training.
63. MetaSym: A Symplectic Meta-learning Framework for Physical Intelligence
Core Problem: Developing scalable and generalizable physics-aware deep learning models that can preserve fundamental physical invariants (like energy and momentum) while adapting efficiently to system heterogeneities with limited data.
Key Innovation: Introduces MetaSym, a novel symplectic meta-learning framework that combines a symplectic inductive bias (from an encoder) with an autoregressive decoder and meta-attention, ensuring preservation of physical invariants and enabling flexible, data-efficient few-shot adaptation to diverse physical systems, outperforming SOTA models.
64. Multi-Level Monte Carlo Training of Neural Operators
Core Problem: Neural operators, used for approximating nonlinear operators related to partial differential equations (PDEs), are expensive to train for large-scale problems at high-resolution, limiting their practical application.
Key Innovation: Presents a Multi-Level Monte Carlo (MLMC) approach to train neural operators by leveraging a hierarchy of resolutions of function discretization, using gradient corrections from fewer samples of fine-resolution data to decrease computational cost while maintaining high accuracy, demonstrating improved efficiency across state-of-the-art models.
65. Moirai 2.0: When Less Is More for Time Series Forecasting
Core Problem: Existing time-series forecasting models, including prior versions of Moirai, can be complex and less efficient, limiting their practical applicability despite good accuracy.
Key Innovation: Moirai 2.0, a simpler decoder-only time-series foundation model, achieves improved probabilistic accuracy and inference efficiency for quantile forecasting and multi-token prediction. It is twice as fast and thirty times smaller than its predecessor while performing better, demonstrating that a simpler architecture can be more effective.
66. Investigating the local heterogeneity of compacted bentonite/sand mixtures
Core Problem: The inevitable heterogeneity of bentonite distribution in compacted bentonite/sand mixtures impacts their hydro-mechanical behavior, which is critical for their function as sealing material in deep geological disposal of radioactive wastes.
Key Innovation: Conducts swelling pressure tests and microstructural observations to show that bentonite distribution heterogeneity is amplified at low bentonite fractions, defining a critical bentonite fraction that explains swelling behavior and the formation of preferential pathways.
67. A double-structure formulation of NCL for saturated bentonite differentiating free and adsorbed water
Core Problem: The normal consolidation line (NCL) of saturated bentonite is not a unique straight line and depends on the initial structure, a phenomenon not considered in existing constitutive models for bentonite.
Key Innovation: Analyzes compression curves to show NCLs differ in the free water-dominated region but coincide in the adsorbed water-dominated region, leading to a proposed double-structure formulation of NCL for saturated bentonite that reflects the coupling effects of two compression mechanisms.
68. Brief communication: Towards disability inclusive risk management
Core Problem: People with disabilities face heightened vulnerability during disasters but remain underrepresented in risk management planning and response.
Key Innovation: A pilot study in Tyrol, Austria, assessed flood exposure and disaster preparedness in facilities serving people with disabilities, revealing significant exposure to flood hazards and critical gaps in risk awareness, preparedness, and inclusive planning, underscoring the urgent need for disability-inclusive disaster risk management.
69. A hybrid physics–AI approach using universal differential equations with state-dependent neural networks for learnable, regionalizable, spatially distributed hydrological modeling
Core Problem: Traditional conceptual hydrological models face structural uncertainties and lack scale-relevant theories, while purely AI methods lack interpretability, making it challenging to achieve both high predictive accuracy for extreme events and explainability in spatially distributed hydrological modeling.
Key Innovation: Introduction of a hybrid physics–AI framework that seamlessly embeds state-dependent neural networks into a spatialized, regionalizable, and fully differentiable process-based hydrological model via universal differential equations (UDEs), enabling end-to-end training with adjoint-based gradients and demonstrating consistently strong performance for streamflow simulations, particularly for flood modeling.
70. Automated stratigraphic interpretation from drillhole lithological descriptions with uncertainty quantification: litho2strat 1.0
Core Problem: Legacy drillhole datasets, while extensive, have limited use for enhancing 3D geological models and understanding subsurface architecture because many lack stratigraphic information, containing only raw lithological descriptions.
Key Innovation: Development of litho2strat 1.0, an open-source methodology combining a search algorithm for geologically plausible stratigraphic orderings and a solution correlation algorithm across multiple drillholes, to automate stratigraphic interpretation from drillhole lithological data, quantify uncertainty, and enhance regional 3D geological modeling.
71. When physics gets in the way: an entropy-based evaluation of conceptual constraints in hybrid hydrological models
Core Problem: The assumption that integrating physics-based conceptual constraints always enhances the performance of hybrid hydrological models is challenged, as the data-driven component might overwrite or circumvent these constraints, leading to questions about the true role and utility of prior physical knowledge.
Key Innovation: Introduction of a quantitative, entropy-based metric from Information Theory to evaluate the relative contributions of physics-based and data-driven components in hybrid hydrological models, revealing that performance often predominantly relies on the data-driven component and that conceptual constraints can sometimes add minimal value or be circumvented by the data-driven part.
72. Detecting the occurrence of preferential flow in soils with stable water isotopes
Core Problem: Identifying preferential flow pathways in soils is challenging with traditional methods, which are often inadequate or labor-intensive.
Key Innovation: Introduced a novel method using vertical soil profiles of stable water isotopes and clustering techniques to identify and assess the locations and variability of preferential flow pathways in soils.
73. Application of Red Relief Image Maps as a complementary tool for mapping volcanic areas: The case of the Monte Amiata volcano, Italy
Core Problem: Identifying and outlining partially hidden volcanic features at intensely vegetated volcanoes is challenging for traditional geological mapping.
Key Innovation: Applied Red Relief Image Maps (RRIMs) derived from high-resolution DTMs, complemented by fieldwork, to effectively identify and outline hidden volcanic features (lava flows, coulees, domes) at the heavily vegetated Monte Amiata volcano, demonstrating its utility for precise mapping in remote or vegetated areas.
74. Windthrow mapping in boreal forests using a spatio-temporal deep learning approach and Sentinel-2 imagery
Core Problem: Accurate and large-scale detection and distinction of windthrow events from other forest disturbances (logging, insect outbreaks) using remote sensing, due to spectral similarities and limitations of pixel-based approaches.
Key Innovation: Development of a spatio-temporal deep learning approach (CResU-Net with CCDC algorithm) using Sentinel-2 imagery for accurate and scalable windthrow detection and characterization across large boreal forest areas, improving timestamp estimation.
75. Moss-biofilm cover modulates hydrodynamic erosion on varied rock-surface morphologies under sheet flow
Core Problem: Unclear understanding of how moss cover interacts with rock micro-topography to regulate runoff, hydraulics, and erosion in rocky desertification areas, particularly under sheet flow.
Key Innovation: Systematically examined the influence of rock surface morphologies and moss cover on runoff, sediment yield, and hydrodynamic parameters using rainfall simulation, demonstrating that moss significantly reduces initial runoff and alters erosion patterns, providing insights for conservation and land management.
76. Identification of fluid-entry clusters and diagnosis of downhole events based on high-frequency water hammer pressure
Core Problem: Existing water hammer pressure-based monitoring methods for hydraulic fracturing are limited to identifying dominant fluid-entry clusters and struggle with fine-scale diagnosis of downhole events.
Key Innovation: A fracturing monitoring method based on high-frequency water hammer pressure was established, employing time-domain, frequency-domain, and composite filtering for signal enhancement. Cepstrum analysis and time-depth conversion enabled identification of multiple fluid-entry clusters and fine-scale diagnosis of downhole events (e.g., diverter effectiveness, plug leakage, slippage), validated by simulated and field data.
77. Transport of colloidal Au-bearing nanoparticles driven by metamorphic decarbonization
Core Problem: The genesis of bonanza-style gold deposits, characterized by ultrahigh-grade Au enrichment, challenges conventional models of chemical transport via aqueous complexes.
Key Innovation: High-pressure experiments and analyses demonstrated that CO2-rich fluids from metamorphic decarbonization create overpressures (>200 MPa), initiating explosive upward migration of Au-Ag nanoparticle-bearing sulfide liquids into porous peridotite. A unified model proposes gas-driven filter pressing and viscous fingering facilitate kilometer-scale transport through lithospheric faults, linking mantle carbon fluxes with crustal mineralization.
78. AVFF-RI: an improved rainfall intensity measurement using common cameras
Core Problem: Rainfall measurements and modeling face critical bottlenecks due to data scarcity and limited performance of conventional methods, hindering precision in urban hydrology and continuous detection with surveillance cameras is computationally intensive.
Key Innovation: AVFF-RI, a novel audio-visual fusion framework leveraging temporally resolved information from spatially distributed surveillance cameras for reliable rainfall estimation, using a two-stage pre-processing and an adaptive fusion stage, demonstrating robustness and better generalization for urban hydrology and disaster response.
79. Budyko scatter reveals interactions between wildfire, land cover change, and climate
Core Problem: A knowledge gap exists regarding how fire-induced changes to the land surface affect carbon and water fluxes over seasonal to decadal time scales, especially with increasing wildfire risk due to climate change.
Key Innovation: Used a land-surface hydrology model and the Budyko framework to demonstrate that wildfire-induced land cover change can significantly reduce annual carbon uptake and shift ecosystems from wetter forests to drier savannas, improving understanding of fire-vegetation dynamics.
80. Co-effects of bedding angle and moisture condition on generalized mode III fracture behavior of sandstone: Insights from acoustic emission localization and 3D scanning reconstruction
Core Problem: Understanding how bedding angle and moisture content co-affect the generalized mode III fracture behavior of anisotropic sandstone is crucial for predicting rock mass stability and related geohazards.
Key Innovation: Systematically analyzed the co-effects of bedding angle and moisture condition on sandstone fracture using acoustic emission and 3D scanning, revealing that both factors significantly influence peak load, stress intensity factor, and energy dissipation, with water weakening rock strength and influencing crack propagation along bedding planes.
81. Persisting Modulation of Interdecadal Pacific Oscillation on Near‐Future Winter Precipitation Projections in Northern Europe
Core Problem: The relative roles of internal variability and external forcing in decadal changes of European winter precipitation remain elusive, leading to uncertainty in near-future projections.
Key Innovation: Identifies the Interdecadal Pacific Oscillation (IPO) as a significant driver of interdecadal changes in winter northern European precipitation, demonstrating its persisting influence into the near future (2015–2050) and showing that accounting for IPO-related influences reduces projection uncertainty by 20-30%.
82. Drying Soil Moisture Dominates Enhancing Summer Soil Moisture‐Temperature Coupling Under Climate Change
Core Problem: The mechanisms and future trends of enhanced soil moisture (SM)-temperature (SM-T) coupling under global warming, which exacerbates extreme climates, are inadequately understood.
Key Innovation: Analysis of multiple reanalysis datasets and ridge regression reveals that drying SM is the dominant contributor to enhanced SM-T coupling, which is projected to further intensify under future emission scenarios due to anthropogenic forcings.
83. Weakening of Regional Hadley Circulation Delays Indian Summer Monsoon Rainfall Onset
Core Problem: The specific causes of delayed Indian Summer Monsoon Rainfall (ISMR) onset, despite its significant impact on agriculture and the economy, remain unclear.
Key Innovation: Shows that delayed ISMR onset is primarily driven by anomalous high pressure over the Arabian Sea and low pressure over the Mascarene High, which weakens the regional Hadley circulation and low-level moist southwesterly winds, offering new insights for improving ISMR onset prediction.
84. Achieving Explainable ENSO Prediction Using Small Data Training
Core Problem: Accurately predicting the spatiotemporal structure of the El Niño–Southern Oscillation (ENSO) remains a persistent challenge for dynamical models, and deep learning models are constrained by climate model biases and lack dynamic interpretability.
Key Innovation: Constructs a novel hybrid model integrating deep learning into a dynamical model, trained on physical-informed small data, achieving unprecedented ENSO prediction skills by improving the representation of leading feedbacks and circumventing climate model biases.
85. naPINN: Noise-Adaptive Physics-Informed Neural Networks for Recovering Physics from Corrupted Measurement
Core Problem: Physics-Informed Neural Networks (PINNs) degrade significantly when recovering physical solutions and discovering governing equations from observational data corrupted by complex measurement noise and gross outliers.
Key Innovation: Noise-Adaptive Physics-Informed Neural Network (naPINN) embeds an energy-based model to learn latent residual distributions, using a trainable reliability gate to adaptively filter high-energy data points and robustly recover physical solutions from severely corrupted measurements.
86. Super-r\'esolution non supervis\'ee d'images hyperspectrales de t\'el\'ed\'etection utilisant un entra\^inement enti\`erement synth\'etique
Core Problem: Most existing hyperspectral single image super-resolution (SISR) methods rely on supervised learning with high-resolution ground truth data, which is often unavailable in practice for hyperspectral remote sensing images.
Key Innovation: An unsupervised learning approach for hyperspectral SISR decomposes the image into endmembers and abundance maps, then trains a neural network to super-resolve these maps using synthetic data generated with the dead leaves model, demonstrating effectiveness and relevance of synthetic data.
87. A General ReLearner: Empowering Spatiotemporal Prediction by Re-learning Input-label Residual
Core Problem: Prevailing spatiotemporal prediction models operate under a unidirectional learning paradigm, leading to suboptimal performance when significant spatiotemporal discrepancies exist between inputs and labels.
Key Innovation: The Spatiotemporal Residual Theorem generalizes spatiotemporal prediction into a bidirectional learning framework, and the ReLearner module augments Spatiotemporal Neural Networks (STNNs) with this capability by relearning input-label spatiotemporal feature residuals, significantly enhancing predictive performance.
88. CAPS: Unifying Attention, Recurrence, and Alignment in Transformer-based Time Series Forecasting
Core Problem: Standard attention mechanisms in time series forecasting entangle temporal structures, and recurrent models sacrifice long-term selection for causal structure.
Key Innovation: CAPS, a structured attention mechanism that decouples global trends, local shocks, and seasonal patterns using SO(2) rotations, three additive gating paths, and a learned Clock mechanism, achieving competitive performance in time series forecasting.
89. Structure-Preserving Learning Improves Geometry Generalization in Neural PDEs
Core Problem: Developing physics foundation models that provide real-time, structure-preserving, and accurate solutions to PDEs, especially under adaptation to unseen geometries.
Key Innovation: General-Geometry Neural Whitney Forms (Geo-NeW), a data-driven finite element method that jointly learns a differential operator and compatible reduced finite element spaces, preserving physical conservation laws and improving generalization to unseen geometries for neural PDEs.
90. Spatiotemporal Decision Transformer for Traffic Coordination
Core Problem: Traffic signal control is challenging, requiring multi-agent coordination to optimize network-wide flow, and existing reinforcement learning methods struggle with coordination and sample efficiency.
Key Innovation: Introduces MADT (Multi-Agent Decision Transformer), reformulating multi-agent traffic signal control as a sequence modeling problem with graph attention and temporal transformer encoder, achieving state-of-the-art performance in reducing travel time and improving coordination.
91. Gromov Wasserstein Optimal Transport for Semantic Correspondences
Core Problem: Establishing accurate and spatially consistent semantic correspondences between image pairs using large foundation models is computationally expensive, and existing methods often rely on combining complementary but costly features.
Key Innovation: Replacing Stable Diffusion features with a Gromov Wasserstein optimal transport algorithm that includes a spatial smoothness prior, significantly boosting DINOv2 baseline performance for semantic correspondence, achieving competitive results with 5-10x greater efficiency.
92. MemCast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning
Core Problem: Existing LLM-based time series forecasters often lack explicit experience accumulation and continual evolution, limiting their long-term effectiveness and adaptability.
Key Innovation: MemCast, a learning-to-memory framework that reformulates time series forecasting as an experience-conditioned reasoning task, organizing learned experience into a hierarchical memory (historical patterns, reasoning wisdom, general laws) and enabling continual evolution via a dynamic confidence adaptation strategy.
93. From Single Scan to Sequential Consistency: A New Paradigm for LIDAR Relocalization
Core Problem: Existing LiDAR relocalization methods struggle in dynamic or ambiguous scenarios because they rely on single-frame inference or neglect spatio-temporal consistency across scans, leading to less robust pose estimation.
Key Innovation: TempLoc, a new LiDAR relocalization framework that enhances robustness by effectively modeling sequential consistency through a Global Coordinate Estimation module, a Prior Coordinate Generation module for inter-frame correspondences, and an Uncertainty-Guided Coordinate Fusion module, yielding more temporally consistent and accurate global 6-DoF poses.
94. Anomaly Detection via Mean Shift Density Enhancement
Core Problem: Existing unsupervised anomaly detection algorithms rarely perform well across different anomaly types and often lack robustness, particularly under noisy settings, excelling only under specific structural assumptions.
Key Innovation: Proposes Mean Shift Density Enhancement (MSDE), a fully unsupervised framework that detects anomalies through their geometric response to density-driven manifold evolution, employing a weighted mean-shift procedure with adaptive, sample-specific density weights, achieving consistently strong, balanced, and robust performance.
95. Causal Graph Learning via Distributional Invariance of Cause-Effect Relationship
Core Problem: Recovering causal graphs from observational data is challenging, and existing methods can be computationally intensive, hindering scalability and efficiency.
Key Innovation: Introduces a new framework for causal graph learning that leverages the distributional invariance of an effect conditioned on its causes across multiple downsampled data subsets, enabling an efficient algorithm with quadratic complexity that uncovers causal relationships with superior or equivalent performance and enhanced scalability.
96. Ultra Fast PDE Solving via Physics Guided Few-step Diffusion
Core Problem: Diffusion-based PDE solvers suffer from high sampling costs and insufficient physical consistency due to many-step iterative sampling and lack of explicit physics constraints.
Key Innovation: Introduction of Phys-Instruct, a physics-guided distillation framework that compresses diffusion PDE solvers into few-step generators for rapid sampling and enhances physical consistency by injecting PDE knowledge through distillation guidance, leading to faster and more accurate PDE solving.
97. SPWOOD: Sparse Partial Weakly-Supervised Oriented Object Detection
Core Problem: Oriented object detection in remote sensing faces high annotation costs due to dense object distribution and variety, and existing methods struggle with sparse or weakly-labeled data.
Key Innovation: Introduction of SPWOOD, the first Sparse Partial Weakly-Supervised Oriented Object Detection framework, featuring a Sparse-annotation-Orientation-and-Scale-aware Student (SOS-Student) model, a Multi-level Pseudo-label Filtering strategy, and a sparse partitioning approach for efficient learning with limited annotations.
98. Decision-oriented benchmarking to transform AI weather forecast access: Application to the Indian monsoon
Core Problem: Current AI weather prediction (AIWP) model evaluations lack decision-oriented frameworks that consider local stakeholders' needs, limiting their utility for populations facing high-impact weather shocks.
Key Innovation: A decision-oriented benchmarking framework connecting meteorology, AI, and social sciences, applied to Indian monsoon forecasting, demonstrating AIWP models' skill in predicting agriculturally relevant onset indices weeks in advance, informing large-scale farmer forecasts.
99. Thermal Comfort Path Planning Tool for Urban Mobility in Austin, Texas
Core Problem: Extreme heat poses a growing challenge for active transportation in cities, and conventional weather reporting fails to capture microclimate variations relevant to pedestrian/cyclist thermal comfort and safety.
Key Innovation: A novel walking and biking route planner that selects paths based on thermal comfort (Universal Thermal Climate Index, UTCI) using high-resolution thermal modeling (SOLWEIG-GPU) and real-time route mapping, identifying 'coolest' routes to reduce heat exposure and related illness.
100. Training and Simulation of Quadrupedal Robot in Adaptive Stair Climbing for Indoor Firefighting: An End-to-End Reinforcement Learning Approach
Core Problem: Quadruped robots for indoor firefighting primary searches face challenges in situational awareness in complex indoor environments and rapid, adaptive stair climbing across diverse staircases.
Key Innovation: Develops a two-stage end-to-end deep reinforcement learning framework that transfers stair-climbing skills from abstract terrain to realistic indoor stair topologies (straight, L-shaped, spiral), enabling unified navigation and locomotion learning for quadruped robots using only local height-map perception.
101. Physics-Based Learning of the Wave Speed Landscape in Complex Media
Core Problem: Conventional reflection imaging fails to recover large-scale variations of the wave velocity landscape in complex media because in vivo measurements are typically limited to reflection geometries.
Key Innovation: A matrix imaging approach that models wave propagation as a trainable multi-layer network, leveraging optimization and deep learning to infer the wave velocity distribution, validated for tumor detection but broadly applicable to any kind of waves and media.
102. Split&Splat: Zero-Shot Panoptic Segmentation via Explicit Instance Modeling and 3D Gaussian Splatting
Core Problem: 3D Gaussian Splatting (GS) lacks object-consistent and semantically aware structure, making it difficult to perform panoptic scene reconstruction and support downstream tasks like object retrieval or 3D editing.
Key Innovation: Split&Splat, a framework for panoptic scene reconstruction using 3DGS that explicitly models object instances by propagating masks across views, reconstructing objects independently, and embedding instance-level semantic descriptors, achieving SOTA performance on segmentation benchmarks.
103. Dataset-Driven Channel Masks in Transformers for Multivariate Time Series
Core Problem: Existing Transformer-based methods for multivariate time series modeling primarily focus on architectural modifications for capturing channel dependency (CD), often neglecting dataset-specific characteristics, which limits their effectiveness.
Key Innovation: Introduces partial channel dependence (PCD) through dataset-driven channel masks (CMs) integrated into Transformer attention matrices, which leverage a similarity matrix and learnable domain parameters to refine channel dependency, improving performance across diverse tasks and datasets.
104. SPAR: Self-supervised Placement-Aware Representation Learning for Distributed Sensing
Core Problem: Existing pretraining methods for distributed sensing are placement-agnostic, failing to account for how sensor placements (spatial locations, structural characteristics) inseparably shape observed signals, limiting robustness and generalization.
Key Innovation: Introduces SPAR, a self-supervised framework that models the duality between signals and positions through spatial and structural positional embeddings and dual reconstruction objectives, treating placement as intrinsic to representation learning, leading to superior robustness and generalization, including for earthquake localization.
105. L2M-Reg: Building-level Uncertainty-aware Registration of Outdoor LiDAR Point Clouds and Semantic 3D City Models
Core Problem: Accurate building-level registration between LiDAR point clouds and semantic 3D city models is challenging due to generalization uncertainty in LoD2 models, hindering urban digital twinning and downstream tasks like change detection.
Key Innovation: Proposes L2M-Reg, a plane-based fine registration method that explicitly accounts for model uncertainty by establishing reliable plane correspondence, building a pseudo-plane-constrained Gauss-Helmert model, and adaptively estimating vertical translation, achieving more accurate and efficient registration.
106. fev-bench: A Realistic Benchmark for Time Series Forecasting
Core Problem: Existing time series forecasting benchmarks have limited domain coverage, overlook real-world settings (e.g., covariates), lack statistical rigor in aggregation, and often lack consistent evaluation infrastructure.
Key Innovation: Proposes fev-bench, a realistic benchmark of 100 forecasting tasks across seven domains, supported by the fev Python library, which employs principled aggregation with bootstrapped confidence intervals to evaluate time series forecasting models.
107. UrbanGraph: Physics-Informed Spatio-Temporal Dynamic Heterogeneous Graphs for Urban Microclimate Prediction
Core Problem: Existing urban microclimate prediction approaches fail to capture physical consistency, spatial dependencies, and temporal variability, leading to inaccurate predictions that impact building energy demand and public health.
Key Innovation: Introduces UrbanGraph, a framework that transforms physical first principles into a dynamic causal topology, explicitly encoding time-varying causalities into a heterogeneous graph structure to ensure physical consistency and data efficiency for urban microclimate prediction.
108. MoGU: Mixture-of-Gaussians with Uncertainty-based Gating for Time Series Forecasting
Core Problem: Traditional Mixture-of-Experts (MoE) frameworks for regression tasks, particularly in volatile time-series forecasting, rely on standard learned gating and often lack reliable, high-fidelity uncertainty quantification.
Key Innovation: Introduces MoGU (Mixture-of-Gaussians with Uncertainty-based Gating), a novel MoE framework that replaces learned gating with an intrinsic routing paradigm where expert-specific uncertainty (predicted variance from Gaussian distributions) dynamically weights expert contributions, improving forecasting accuracy and providing efficient prediction intervals.
109. UniADC: A Unified Framework for Anomaly Detection and Classification
Core Problem: Existing methods for anomaly detection and classification treat these tasks separately, neglecting their inherent correlations and limiting information sharing, and struggle with severe imbalance between normal and anomalous pixel distributions, especially with few or no anomaly images.
Key Innovation: UniADC, a unified framework, simultaneously detects anomalous regions and identifies their categories using a training-free Controllable Inpainting Network (for synthesizing/augmenting anomaly data) and an Implicit-Normal Discriminator (for implicitly modeling the normal state), outperforming existing methods in detection, localization, and classification.
110. Spectral Text Fusion: A Frequency-Aware Approach to Multimodal Time-Series Forecasting
Core Problem: Existing multimodal time series forecasting methods neglect the multiscale temporal influences of contextual information (like time-series cycles and dynamic shifts) by only performing local alignment of textual features with time-series patterns.
Key Innovation: SpecTF, a frequency-aware framework, integrates textual data into time series forecasting by projecting textual embeddings into the frequency domain and fusing them with time series' spectral components using a lightweight cross-attention mechanism, adaptively reweighting frequency bands based on textual relevance.
111. COMET: Codebook-based Online-adaptive Multi-scale Embedding for Time-series Anomaly Detection
Core Problem: Time series anomaly detection struggles with capturing temporal dependencies and multivariate correlations within patch-level representation learning, is limited by single-scale pattern analysis, and is vulnerable to distribution shifts at inference time.
Key Innovation: COMET introduces Multi-scale Patch Encoding, a Vector-Quantized Coreset for learning representative normal patterns and dual-score anomaly detection, and Online Codebook Adaptation for dynamic model adaptation at inference, achieving state-of-the-art performance across diverse environments.
112. Cross-Modal Alignment and Fusion for RGB-D Transmission-Line Defect Detection
Core Problem: Automated UAV inspection for transmission line defect detection struggles with small-scale defects, complex backgrounds, and illumination variations, especially for geometrically subtle defects under limited chromatic contrast.
Key Innovation: CMAFNet integrates RGB appearance and depth geometry through a purify-then-fuse paradigm. It uses a Semantic Recomposition Module for dictionary-based feature purification and a Contextual Semantic Integration Framework with partial-channel attention, enforcing cross-modal alignment via position-wise normalization for robust defect detection.
113. SurfSplat: Conquering Feedforward 2D Gaussian Splatting with Surface Continuity Priors
Core Problem: Reconstructing 3D scenes from sparse images often yields discrete, color-biased point clouds with severe artifacts under close-up views, failing to produce continuous surfaces with accurate geometry and texture.
Key Innovation: SurfSplat, a feedforward framework based on 2D Gaussian Splatting, incorporates a surface continuity prior and a forced alpha blending strategy to reconstruct coherent geometry and faithful textures. It consistently outperforms prior methods on 3D reconstruction quality, especially at high resolution.
114. SEDformer: Event-Synchronous Spiking Transformers for Irregular Telemetry Time Series Forecasting
Core Problem: Existing Graph- and Transformer-based forecasters for irregular multivariate time series (IMTS) ignore its Sparsity-Event Duality (SED) property, leading to inflated sequences, computation at non-informative steps, and weakened event semantics.
Key Innovation: SEDformer, an SED-enhanced Spiking Transformer, uses a SED-based Spike Encoder, an Event-Preserving Temporal Downsampling module, and SED-based Spike Transformer blocks to faithfully model IMTS. It achieves state-of-the-art forecasting accuracy while reducing energy and memory usage by aligning with the event-driven nature of IMTS.
115. A data-driven comparative risk assessment of marine traffic accidents using an object-oriented Bayesian network
Core Problem: Expanding Arctic shipping faces increased risks from rapidly changing ice regimes, severe weather, and limited infrastructure, necessitating a robust, accident-type-specific risk assessment framework for marine traffic accidents.
Key Innovation: An object-oriented Bayesian network (OOBN) is developed for comparative risk assessment of marine traffic accidents (collisions, groundings, machinery damages) in the Arctic, integrating human, technical, organizational, and environmental factors, and revealing interaction-driven risk amplification from environmental stressors.
116. A novel failure mode and effects analysis model enhanced with systems theory and artificial intelligence for dynamic positioning systems in offshore operations
Core Problem: Traditional FMEA struggles to reliably analyze the complex and uncertain failure modes of dynamic positioning (DP) systems in offshore operations, which can lead to severe accidents.
Key Innovation: A novel T-spherical fuzzy FMEA model is proposed, integrating systems theory and AI techniques for enhanced risk perception, data fusion, and robust ranking, effectively capturing system-level hidden failure modes and identifying systemic and control loop failures as primary risks in DP systems.
117. Investigation on the characteristics of wave propagation in reef lagoon based on the particle method
Core Problem: Understanding wave propagation characteristics in complex reef lagoon topographies, especially large free-surface deformation from wave breaking, is challenging for computational models.
Key Innovation: Developed a numerical algorithm based on the meshless particle method to simulate wave propagation and breaking in reef lagoons, providing equations for wave height and energy attenuation coefficients, which are crucial for understanding coastal dynamics.
118. Experimental assessment and prediction of wave loading around abrupt depth transitions
Core Problem: Abrupt depth transitions increase wave non-linearity and the likelihood of extreme events, making it difficult to identify critical locations for wave loading on structures using free field properties alone.
Key Innovation: Conducted a comprehensive experimental analysis of wave loading on a vertical cylinder around a shoal bathymetry, demonstrating that extreme loads are most frequent above the crest and are influenced by wave breaking, and investigated Morison's equation for prediction.
119. Soil moisture monitoring with cosmogenic neutrons: an asset for the development and assessment of soil moisture products in the state of Brandenburg (Germany)
Core Problem: The German federal state of Brandenburg has been particularly impacted by soil moisture droughts, requiring timely and informed management of water-related risks and improved assessment of soil moisture products.
Key Innovation: Introduced a novel soil moisture monitoring network based on cosmic-ray neutron sensing (CRNS) technology across eight sites in Brandenburg, providing openly accessible data and demonstrating its utility to evaluate and improve large-scale soil moisture products (ERA5-Land, SWI, C3S, SWAP model) for water-related risk management.
120. The Western United States Large Forest-Fire Stochastic Simulator (WULFFSS) 1.0: a monthly gridded forest-fire model using interpretable statistics
Core Problem: Accurately simulating large forest fires in the western US at a monthly gridded resolution, capturing interannual variability and regional differences in fire frequency and size, while maintaining interpretability and efficient ensemble generation.
Key Innovation: Development of WULFFSS 1.0, a stochastic monthly gridded forest-fire model for the western US (12 km resolution) that calculates fire probabilities and sizes using interpretable multiple logistic and linear regressions, forced by vegetation, topographic, anthropogenic, and climate factors, and demonstrating strong performance in capturing observed fire patterns and interannual variability.
121. GeoDS (v.1.0): a simple Geographical DownScaling model for long-term precipitation data over complex terrains
Core Problem: Global climate models lack the spatial resolution needed for fine-scale precipitation data over complex terrains in long-term (millennial or longer) simulations, and existing downscaling techniques often have limitations for such periods.
Key Innovation: Presentation of GeoDS (v.1.0), a simple, computationally inexpensive, topography-based model that downscales precipitation fields in complex areas by computing a topographic exposure index from large-scale winds and terrain configuration, demonstrating its ability to capture fine-scale precipitation patterns and robustness for paleoclimate applications.
122. A probabilistic approach to wildfire spread prediction using a denoising diffusion surrogate model
Core Problem: Traditional deterministic approaches to wildfire spread prediction struggle to capture the inherent uncertainty and variability of wildfire dynamics, limiting their ability to provide probabilistic risk information for assessment and operational planning.
Key Innovation: Proposal of a stochastic framework for wildfire spread prediction using deep generative denoising diffusion models with ensemble sampling, which generates probabilistic forecasts by sampling multiple plausible future scenarios, outperforming deterministic baselines in accuracy, spatial coherence, distributional quality, and probabilistic capability (hit rate).
123. BuRNN (v1.0): a data-driven fire model
Core Problem: Accurately simulating burned area globally and understanding past fire behavior remains challenging with traditional numerical models, and existing process-based fire models often have limitations in performance.
Key Innovation: Presentation of BuRNN (v1.0), a data-driven model using Long Short-Term Memory networks to simulate burned area on a global 0.5° × 0.5° grid with monthly resolution, outperforming process-based fire models and providing insights into regional drivers of burned area through explainable AI.
124. Direct assimilation of ground-based microwave radiometer observations with machine learning bias correction based on developments of RTTOV-gb v1.0 and WRFDA v4.5
Core Problem: The application of ground-based microwave radiometers (GMWRs) for atmospheric observations has traditionally focused on indirect assimilation of retrieved profiles, and direct assimilation of radiances, along with effective bias correction, is needed to further improve initial conditions and forecasts.
Key Innovation: Development of a direct assimilation capability for GMWR radiance observations within the WRFDA system (v4.5), coupled with a random forest-based machine learning bias correction scheme, which significantly reduces biases and standard deviations, leading to improved initial conditions and sustained benefits for temperature, humidity, and precipitation forecasts.
125. Revealing the causes of groundwater level dynamics in seasonally frozen soil zones using interpretable deep learning models
Core Problem: Accurately characterizing and understanding the underlying causes of groundwater level dynamics in seasonal frozen soil regions.
Key Innovation: Proposed an interpretable deep learning method (LSTM with Expected Gradients) to simulate and reveal dominant factors and mechanisms of groundwater level dynamics in seasonally frozen soil, identifying specific variation types and the regulatory role of frozen-thaw processes.
126. Bedrock geology controls on new water fractions and catchment functioning in contrasted nested catchments
Core Problem: Lack of substantial understanding on how landscape characteristics (bedrock geology, land cover) shape water storage and release at the catchment scale.
Key Innovation: Used 13 years of stable isotope measurements and hydrometeorological data to link water storage and release functions to bedrock geology and land cover, revealing how bedrock permeability influences runoff coefficients and the fraction of new water in streamflow generation.
127. Modelling of groundwater salt pollution in semi-arid watershed disturbed by agricultural activities: Lake Tuz (Salt Lake) Basin, Turkey
Core Problem: The need to model and investigate groundwater salt pollution/contamination in the Lake Tuz Basin, a semi-arid watershed with intensive agricultural activities and natural salty lakes, to understand and predict salinity changes.
Key Innovation: Developed and calibrated a spatiotemporal groundwater quality (salinity) model using MT3DMS linked with a modified Clonal Selection Algorithm, simulating groundwater salt pollution for 19 years and predicting future concentrations based on anthropogenic and natural factors in the Lake Tuz watershed.
128. Effect of fracture shear dilation on flow anisotropy for variable normal stress and fracture size
Core Problem: Fracture shear can induce flow channeling and anisotropy, which is crucial for geothermal sites, but its impact under variable normal stress, shear displacement, and fracture size is not fully understood.
Key Innovation: A numerical shear model simulating asperity degradation was combined with flow simulations to examine flow anisotropy. Results showed significantly increased permeability anisotropy ratio with higher normal stress and that anisotropy remained evident and significant even with increasing fracture size. Perpendicular flow was enhanced at both laboratory and reservoir scales.
129. Experimental and 3D numerical study on the damage and fracture mechanisms of hot dry rock cores under microwave irradiation
Core Problem: Hot dry rock (HDR) reservoirs require stimulation to improve permeability, and microwave irradiation is a potential low-energy approach, but its damage and fracture mechanisms are not fully characterized.
Key Innovation: Laboratory coaxial microwave irradiation experiments on HDR cores revealed temperature rise responses and typical fracture patterns. A 3D multiphysics numerical simulation method (COMSOL + peridynamics) was developed, accurately capturing temperature evolution, damage, and fracture processes, revealing temperature-field-dominated damage initiation and directional propagation.
130. Influence of laser parameters on rock damage: an experimental exploration and machine learning-based predictive modeling
Core Problem: A comprehensive understanding of how laser parameters influence rock-breaking effects, particularly changes in rock strength after laser irradiation, and the development of predictive optimization models is lacking.
Key Innovation: Systematic experiments analyzed the effects of laser power, irradiation time, spot shape size, and irradiation distance on granite samples. A predictive optimization model using BPNN and NSGA-II was developed, identifying the significance order of parameters (P > t > S > d) and an optimal combination for rock-breaking efficiency and strength decay.
131. Constructing large-scale high-fidelity fracture networks based on generative AI
Core Problem: Accurate, large-scale, high-fidelity fracture network modeling is crucial for various fields, but field data typically provides only small-scale images, which are insufficient for engineering-scale analysis.
Key Innovation: A novel generative AI algorithm, Upscaling-GAN, was introduced to generate large-scale high-fidelity fracture networks by learning from small-scale images. This two-stage process accurately characterizes geometrical and topological structures, captures spatial variability of fracture apertures without laborious preprocessing, and maintains low GPU memory consumption for large images.
132. Rapidly improving the acid-fracture conductivity in deep and ultra-deep carbonate reservoirs through mineral alteration: a new method
Core Problem: High closure stress and acid-induced damage lead to fracture closure and conductivity degradation in deep carbonate reservoirs, hindering efficient resource development, and existing mineral alteration processes are too slow.
Key Innovation: A new method using Na2HPO4 + H3PO4 buffer solution (PPN) at 200 °C rapidly enhanced fracture conductivity in dense carbonate rocks. After 4 hours, fracture conductivity increased by factors of 29.4 and 19.0, respectively, for Mao-kou and Jialingjiang formations, by converting carbonate minerals into harder hydroxyapatite and repairing acid-induced damage, thus enhancing rock strength and deformation resistance.
133. Extreme precipitation in eastern China: a centennial-scale analysis across multiple river basins based on return period thresholds
Core Problem: Traditional methods for identifying extreme precipitation based on percentile thresholds (PTs) are insufficient in sensitivity to changes in precipitation extremes under global warming.
Key Innovation: Introduction and verification of "return period thresholds" (RPT) from hydrology for extreme precipitation definition, demonstrating their superior sensitivity in reflecting actual changes, and projecting upward trends in thresholds across all basins by 2100, with implications for enhanced monitoring.
134. Effect of gradation characteristics on seepage failure in foam-conditioned gap-graded soils for EPB shield tunneling
Core Problem: Muck spewing at the screw conveyor outlet in water-rich coarse-grained soils during shield tunneling poses significant risks, and the influence of soil gradation on seepage failure in foam-conditioned soils is not well understood.
Key Innovation: Systematically investigated the influence of gradation characteristics (fines content, gap ratio) and foam injection on seepage failure modes and hydraulic responses in foam-conditioned gap-graded soils, revealing that pore structure and fines content control foam stability and mitigate erosion.
135. Warming of the Mid‐Troposphere Driven by Evapotranspiration During Compound Heatwave and Drought Events
Core Problem: The vertical transport height of evapotranspired water vapor (EWV) and its warming effect during compound heatwave and drought events (CHDE) remain poorly understood, limiting forecasting capabilities.
Key Innovation: Uses multi-source atmospheric water vapor isotope datasets to identify that EWV can reach altitudes up to 500 hPa and plays a key role in warming the mid-troposphere during CHDE, suggesting improved forecasting capabilities for extreme events.
136. Impact of Inter‐Basin Interactions on ENSO‐Associated Hadley Circulation Adjustments
Core Problem: The relative importance and underlying processes of pure atmospheric and coupled ocean-atmosphere pathways in driving remote Hadley Circulation (HC) adjustments across the Atlantic and Indian Oceans during El Niño events remain poorly understood.
Key Innovation: Integrates observations with climate model experiments to disentangle the roles of pure atmospheric and coupled ocean-atmosphere pathways, establishing that the pure atmospheric pathway is the dominant driver of ENSO-related tropical circulation adjustments.
137. The Dominant Role of the Electron Isotropy Boundary in Controlling Earth's Outer Radiation Belt Electron Lifetimes
Core Problem: The direct and quantitative impact of field-line curvature scattering (FLCS) on controlling outer radiation belt electron lifetimes has never been directly assessed, despite its believed role in forming electron isotropy boundaries.
Key Innovation: Provides the first direct and quantitative evidence, combining simultaneous satellite observations and simulations, that FLCS-induced electron loss outside the electron isotropy boundary dominantly controls Earth's outer radiation belt electron lifetimes.
138. Dawn‐Side Anomaly in Sudden Geomagnetic Field Responses to Solar Wind Pressure Discontinuities During the 10 May and 10 October 2024 Geomagnetic Storms
Core Problem: Anomalous dawn-side sudden geomagnetic field responses (SCs) of opposite polarity during intense geomagnetic storms deviate from typical low-latitude SCs, and their source is not fully understood.
Key Innovation: Examines SC responses during two major geomagnetic storms, consistently revealing anomalous dawn-side low-latitude SCs, and suggests the main impulse of the Disturbance Polar (DP) field extending equatorward as the most likely source, providing new insights into magnetosphere-ionosphere coupling.
139. Tropical Rainforests Exhibit Higher Multidimensional Stability Than Rubber Plantations Despite Lower Resistance to Temperature Anomalies
Core Problem: The impact of rapid rubber plantation expansion on regional ecosystem stability compared to diverse tropical rainforests, particularly concerning their response to climate anomalies, is not well understood.
Key Innovation: Compares multidimensional stability metrics (temporal stability, resilience, resistance) between tropical rainforests and rubber plantations using 37 years of satellite data, revealing that rainforests exhibit higher overall stability and resilience, and highlighting threats of conversion to regional ecosystem stability.
140. Uncovering the Generation Mechanism of Low‐Frequency Chorus Waves (<0.1 fce_eq ${\boldsymbol{f}}_{\mathbf{c}\mathbf{e}\mathbf{\_}\mathbf{e}\mathbf{q}}$) During Active Geomagnetic Environments
Core Problem: The generation mechanism and environmental drivers of low-frequency chorus waves (below 0.1 fce_eq), which have distinct effects on radiation belt dynamics, remain poorly understood.
Key Innovation: Uses Van Allen Probes data to elucidate the generation mechanism of low-frequency chorus waves, showing they are excited by the coexistence of isotropic low-energy and anisotropic high-energy electrons during concurrent magnetic storms and substorms.
141. Incident-Guided Spatiotemporal Traffic Forecasting
Core Problem: Existing spatiotemporal traffic forecasting models overlook the substantial impact of sudden incidents (e.g., traffic accidents, adverse weather) on temporal patterns, hindering prediction accuracy.
Key Innovation: The Incident-Guided Spatiotemporal Graph Neural Network (IGSTGNN) explicitly models incident impact through an Incident-Context Spatial Fusion (ICSF) module and a Temporal Incident Impact Decay (TIID) module, achieving state-of-the-art performance on a new incident-aligned traffic dataset.
142. EEO-TFV: Escape-Explore Optimizer for Web-Scale Time-Series Forecasting and Vision Analysis
Core Problem: Transformer-based foundation models for time-series forecasting and image segmentation suffer from error accumulation in long-sequence prediction and vulnerability to out-of-distribution samples, exacerbated by optimization difficulties in high-dimensional Web-scale data.
Key Innovation: A lightweight Transformer architecture combined with a novel Escape-Explore Optimizer (EEO) enhances exploration and generalization while avoiding sharp minima and saddle-point traps, achieving state-of-the-art performance on 11 time-series datasets and medical image segmentation.
143. Copula-Based Aggregation and Context-Aware Conformal Prediction for Reliable Renewable Energy Forecasting
Core Problem: Constructing coherent and calibrated fleet-level probabilistic forecasts for renewable energy from heterogeneous site-level inputs is challenging due to complex cross-site dependencies and aggregation-induced miscalibration.
Key Innovation: Proposes a calibrated probabilistic aggregation framework that integrates copula-based dependence modeling to capture cross-site correlations with Context-Aware Conformal Prediction (CACP) to correct miscalibration, achieving reliable fleet-level forecasts with valid coverage and sharp prediction intervals.
144. Learnable Koopman-Enhanced Transformer-Based Time Series Forecasting with Spectral Control
Core Problem: Improving the stability, interpretability, and performance of deep learning models for time series forecasting by integrating principles from linear dynamical systems.
Key Innovation: Introduces a unified family of learnable Koopman operator parameterizations that integrate linear dynamical systems theory with transformer-based architectures, enabling explicit control over spectral properties and achieving improved bias-variance trade-off, conditioning, and interpretability in time series forecasting.
145. LmPT: Conditional Point Transformer for Anatomical Landmark Detection on 3D Point Clouds
Core Problem: Accurate and automatic identification of anatomical landmarks on 3D point clouds is challenging due to time-consuming manual processes, inter-observer variability, and limitations of rule-based methods.
Key Innovation: Landmark Point Transformer (LmPT), a method for automatic anatomical landmark detection on point clouds that incorporates a conditioning mechanism for cross-species learning and demonstrates generalization across different species.
146. Koopman Autoencoders with Continuous-Time Latent Dynamics for Fluid Dynamics Forecasting
Core Problem: Existing Koopman autoencoders for fluid dynamics forecasting operate in discrete-time, limiting temporal flexibility and generalization beyond training regimes, leading to trade-offs between short-term accuracy and long-horizon stability.
Key Innovation: Introduces a continuous-time Koopman framework that models latent evolution through numerical integration, demonstrating robustness to temporal resolution, generalization beyond training, and efficient long-horizon forecasting for fluid dynamics.
147. Variational Sparse Paired Autoencoders (vsPAIR) for Inverse Problems and Uncertainty Quantification
Core Problem: Many scientific and engineering inverse problems require not just point estimates but also interpretable uncertainty quantification, with fast inference, which remains challenging.
Key Innovation: Proposes Variational Sparse Paired Autoencoder (vsPAIR), an architecture that pairs a VAE for observations with a sparse VAE for quantities of interest, connected by a latent mapping. It enables uncertainty estimation, interpretability through sparse encodings, and fast inference for inverse problems like blind inpainting and computed tomography.
148. Function-Space Empirical Bayes Regularisation with Large Vision-Language Model Priors
Core Problem: Bayesian deep learning struggles with designing informative prior distributions that scale effectively to high-dimensional data, as existing functional variational inference methods often rely on limited Gaussian Process priors.
Key Innovation: VLM-FS-EB, a novel function-space empirical Bayes regularisation framework that leverages large vision-language models (VLMs) to generate semantically meaningful context points and construct expressive functional priors, improving predictive performance and uncertainty estimates, especially for OOD detection and data-scarce regimes.
149. Beyond Cropping and Rotation: Automated Evolution of Powerful Task-Specific Augmentations with Generative Models
Core Problem: While generative models offer diverse and realistic data augmentations, poorly matched augmentations can degrade model performance, and existing methods for automated augmentation design are limited to traditional transformations.
Key Innovation: EvoAug, an automated augmentation learning pipeline that leverages generative models (conditional diffusion, few-shot NeRFs) and an evolutionary algorithm to learn optimal task-specific augmentations, introducing stochastic augmentation trees for structured transformations and demonstrating strong performance in low-data settings.
150. FSOD-VFM: Few-Shot Object Detection with Vision Foundation Models and Graph Diffusion
Core Problem: Existing few-shot object detection methods using vision foundation models often produce fragmented bounding boxes and numerous false-positive proposals, leading to inaccurate object detections.
Key Innovation: FSOD-VFM, a framework integrating a universal proposal network, SAM2, and DINOv2 features, enhanced by a novel graph-based confidence reweighting method that refines proposal scores to improve detection granularity and reduce false positives.
151. FARTrack: Fast Autoregressive Visual Tracking with High Performance
Core Problem: High-performance visual trackers often suffer from slow inference speeds, making them impractical for deployment on resource-constrained devices.
Key Innovation: FARTrack, a Fast Auto-Regressive Tracking framework that achieves high performance and efficient execution by introducing Task-Specific Self-Distillation for model compression and Inter-frame Autoregressive Sparsification for redundant-to-essential token optimization, enabling real-time tracking with competitive accuracy.
152. EventFlash: Towards Efficient MLLMs for Event-Based Vision
Core Problem: Current event-based multimodal large language models (MLLMs) for robust perception are computationally costly due to dense image-like processing, overlooking the spatiotemporal sparsity of event streams.
Key Innovation: EventFlash, a novel and efficient MLLM that explores spatiotemporal token sparsification to reduce data redundancy and accelerate inference, achieved through an adaptive temporal window aggregation module and a sparse density-guided attention module, supported by the large-scale EventMind dataset.
153. LEVIO: Lightweight Embedded Visual Inertial Odometry for Resource-Constrained Devices
Core Problem: State-of-the-art visual-inertial odometry (VIO) systems are too computationally demanding for resource-constrained hardware, such as micro-drones and smart glasses, limiting their use in mobile robotics and augmented reality (AR) applications.
Key Innovation: Presents LEVIO, a fully featured VIO pipeline optimized for ultra-low-power compute platforms, incorporating established VIO components with a computationally efficient architecture emphasizing parallelization and low memory usage, achieving 6-DoF real-time sensing at low power consumption.
154. Z3D: Zero-Shot 3D Visual Grounding from Images
Core Problem: Prior zero-shot 3D visual grounding methods suffer from significant performance degradation due to bottlenecks in generating high-quality 3D bounding box proposals and advanced reasoning.
Key Innovation: Introduces Z3D, a universal grounding pipeline that flexibly operates on multi-view images, addressing bottlenecks with a state-of-the-art zero-shot 3D instance segmentation method and advanced prompt-based reasoning via modern VLMs, achieving state-of-the-art performance.
155. Multi-Resolution Alignment for Voxel Sparsity in Camera-Based 3D Semantic Scene Completion
Core Problem: Camera-based 3D semantic scene completion methods for autonomous driving face challenges from voxel sparsity, where a large portion of voxels are empty, limiting optimization efficiency and model performance.
Key Innovation: Proposes a Multi-Resolution Alignment (MRA) approach that mitigates voxel sparsity by exploiting scene and instance-level alignment across multi-resolution 3D features as auxiliary supervision, using a Multi-resolution View Transformer, Cubic Semantic Anisotropy, and Critical Distribution Alignment.
156. TIPS Over Tricks: Simple Prompts for Effective Zero-shot Anomaly Detection
Core Problem: Zero-shot anomaly detection using Vision-Language Models (VLMs) like CLIP suffers from coarse image-text alignment, leading to spatial misalignment and weak sensitivity to fine-grained anomalies, often requiring complex auxiliary modules.
Key Innovation: Revisiting the VLM backbone by using TIPS (trained with spatially aware objectives) and addressing the global-local feature gap with decoupled prompts and local evidence injection, improving image-level and pixel-level performance across industrial datasets.
157. CTTVAE: Latent Space Structuring for Conditional Tabular Data Generation on Imbalanced Datasets
Core Problem: Generating synthetic tabular data for imbalanced datasets, especially for rare, high-impact events, is challenging as most generative models overlook minority groups or produce unhelpful samples for downstream learning.
Key Innovation: Introduction of CTTVAE, a Conditional Transformer-based Tabular Variational Autoencoder with a class-aware triplet margin loss for latent space structuring and a training-by-sampling strategy for adaptive exposure to underrepresented groups, yielding more representative and utility-aligned samples for minority classes.
158. Referring Industrial Anomaly Segmentation
Core Problem: Traditional Industrial Anomaly Detection (IAD) methods suffer from rough localizations, overfitting due to scarce/imbalanced data, and the 'One Anomaly Class, One Model' limitation.
Key Innovation: Proposal of Referring Industrial Anomaly Segmentation (RIAS), a language-guided paradigm for precise anomaly mask generation and diverse anomaly detection with a single model, supported by the MVTec-Ref dataset and the DQFormer benchmark with Language-Gated Multi-Level Aggregation.
159. QuAIL: Quality-Aware Inertial Learning for Robust Training under Data Corruption
Core Problem: Tabular ML systems are often trained on data with non-uniform corruption (noise, missing entries, biases), but reliability indicators are typically column-level, limiting existing robustness techniques.
Key Innovation: Introduction of QuAIL, a quality-informed training mechanism that incorporates feature reliability priors directly into learning via a learnable feature-modulation layer and a quality-dependent proximal regularizer, stabilizing optimization under structured corruption and improving performance.
160. LLM-Inspired Pretrain-Then-Finetune for Small-Data, Large-Scale Optimization
Core Problem: Small-data, large-scale decision problems where firms must make many simultaneous operational decisions with few, noisy data points per instance, requiring methods to leverage domain knowledge and synthetic data.
Key Innovation: A pretrain-then-finetune approach using a designed Transformer model, pretrained on large-scale, domain-informed synthetic data and fine-tuned on real observations, to inject domain knowledge, train high-capacity models with scarce real data, and improve alignment with the true data-generating regime.
161. Edge-Optimized Vision-Language Models for Underground Infrastructure Assessment
Core Problem: Automated detection and human-readable summarization of structural deficiencies in underground infrastructure is challenging, especially on resource-constrained edge devices.
Key Innovation: A two-stage edge-optimized pipeline combining a lightweight segmentation model (RAPID-SCAN) with a fine-tuned Vision-Language Model (Phi-3.5 VLM) for efficient defect detection and natural language summarization, deployed on a mobile robotic platform.
162. Downscaling land surface temperature data using edge detection and block-diagonal Gaussian process regression
Core Problem: Accurate and high-resolution estimation of land surface temperature (LST) from satellite data (e.g., ECOSTRESS) is challenging due to resolution limitations, especially when needing to capture fine-scale structures like agricultural fields.
Key Innovation: Develops a novel statistical method for LST downscaling using edge detection from Landsat 8 data to identify agricultural field boundaries, and a block-diagonal Gaussian process (BDGP) model to capture spatial structure and provide high-resolution LST estimates with uncertainty quantification.
163. IMAGINE: Intelligent Multi-Agent Godot-based Indoor Networked Exploration
Core Problem: Autonomous exploration of unknown, GNSS-denied indoor environments by collaborative UAVs faces significant challenges in coordination, perception, and decentralized decision-making, often limited by discrete actions, centralized formulations, and connectivity assumptions.
Key Innovation: Implements IMAGINE, a Multi-Agent Reinforcement Learning (MARL) framework in a high-fidelity Godot simulation for UAVs, enabling emergent collaborative behaviors, continuous action spaces, and robust exploration under communication constraints, addressing several limitations of prior research.
164. Pi-GS: Sparse-View Gaussian Splatting with Dense {\pi}^3 Initialization
Core Problem: 3D Gaussian Splatting (3DGS) relies heavily on accurate camera poses and high-quality point cloud initialization, which are difficult to obtain in sparse-view scenarios where traditional Structure from Motion (SfM) pipelines often fail.
Key Innovation: A robust method utilizing {\pi}^3, a reference-free point cloud estimation network, integrated with dense initialization and a regularization scheme (uncertainty-guided depth supervision, normal consistency loss, depth warping) to mitigate geometric inaccuracies and achieve state-of-the-art performance in sparse-view Gaussian Splatting.
165. Simulation-Based Inference via Regression Projection and Batched Discrepancies
Core Problem: Simulation-based inference methods can be computationally intensive, and their identifiability limitations when using low-information summaries are not always clear.
Key Innovation: A lightweight simulation-based inference method using regression-based projection and batched discrepancies, providing a simple, parallelizable pseudo-posterior, with consistency proofs and characterization of asymptotic concentration and identifiability.
166. Contextual Causal Bayesian Optimisation
Core Problem: Existing Causal Bayesian Optimization and Contextual Bayesian Optimization approaches are distinct and have limitations in scenarios requiring joint optimization over policies and their variable definitions, leading to suboptimal results.
Key Innovation: A unified framework for contextual and causal Bayesian optimization, proposing a novel algorithm that jointly optimizes policies and their variable sets, extending and unifying previous approaches, and deriving worst-case and instance-dependent high-probability regret bounds.
167. Object Fidelity Diffusion for Remote Sensing Image Generation
Core Problem: Existing diffusion models for remote sensing image generation produce low-fidelity images, particularly for morphological details, which negatively impacts the robustness and reliability of object detection models.
Key Innovation: Proposes Object Fidelity Diffusion (OF-Diff), which extracts prior object shapes based on layout, uses a dual-branch diffusion model with diffusion consistency loss, and fine-tunes with DDPO to generate high-fidelity, diverse, and semantically consistent remote sensing images.
168. Material-informed Gaussian Splatting for 3D World Reconstruction in a Digital Twin
Core Problem: 3D reconstruction for Digital Twins often relies on LiDAR, which provides accurate geometry but lacks semantics and textures, and struggles with certain materials (e.g., glass). Traditional LiDAR-camera fusion is complex.
Key Innovation: A camera-only pipeline that reconstructs scenes using 3D Gaussian Splatting, extracts semantic material masks, converts Gaussian representations to mesh surfaces with projected material labels, and assigns physics-based material properties. This provides photorealistic reconstruction with physics-based material assignment, comparable to LiDAR-camera fusion but without hardware complexity.
169. CountZES: Counting via Zero-Shot Exemplar Selection
Core Problem: Zero-shot object counting in complex scenes is challenging because existing methods often rely on off-the-shelf open-vocabulary detectors that suffer from noise and multi-instance proposals, or random patch sampling that fails to delineate instances accurately.
Key Innovation: CountZES, an inference-only approach, discovers diverse exemplars for zero-shot object counting through three synergistic stages: Detection-Anchored Exemplar (refines OVD detections), Density-Guided Exemplar (identifies statistically consistent exemplars), and Feature-Consensus Exemplar (reinforces visual coherence), achieving superior performance and effective generalization.
170. Rectification Reimagined: A Unified Mamba Model for Image Correction and Rectangling with Prompts
Core Problem: Image correction and rectangling tasks in photography systems often rely on task-specific architectures, limiting generalization and effective application across diverse distortions.
Key Innovation: UniRect, a unified Mamba-based framework, addresses image correction and rectangling from a consistent distortion rectification perspective. It uses a dual-component structure (Residual Progressive Thin-Plate Spline Deformation Module and Residual Mamba Blocks Restoration Module) and a Sparse Mixture-of-Experts structure to handle diverse distortions, achieving state-of-the-art performance.
171. Comprehensive Machine Learning Benchmarking for Fringe Projection Profilometry with Photorealistic Synthetic Data
Core Problem: Machine learning approaches for Fringe Projection Profilometry (FPP) are hampered by a lack of large, diverse datasets and standardized benchmarking protocols, making it difficult to evaluate and advance single-shot FPP for 3D depth prediction.
Key Innovation: Introduces the first open-source, photorealistic synthetic dataset for FPP (15,600 fringe images, 300 depth reconstructions) and conducts comprehensive benchmarking. It identifies optimal learning configurations, demonstrates the importance of individual depth normalization and background fringes, and evaluates architectures, revealing that information deficit, not model design, is the primary limitation for single-shot FPP accuracy.
172. DeepUrban: Interaction-Aware Trajectory Prediction and Planning for Automated Driving by Aerial Imagery
Core Problem: Autonomous driving benchmarks lack scenarios with dense traffic, which is crucial for robust prediction and planning capabilities and understanding complex interactions among road users.
Key Innovation: DeepUrban, a new drone dataset providing a rich collection of 3D traffic objects extracted from high-resolution aerial images, enhances trajectory prediction and planning benchmarks, boosting accuracy of vehicle predictions and planning by up to 44.3% on ADE/FDE metrics.
173. Model Optimization for Multi-Camera 3D Detection and Tracking
Core Problem: Optimizing multi-camera 3D detection and tracking frameworks like Sparse4D for efficiency and robustness under varying conditions (e.g., reduced frame rates, quantization, mixed-precision) while maintaining identity stability is challenging.
Key Innovation: Evaluation of Sparse4D under various optimization strategies reveals stability under moderate FPS reductions, identifies optimal selective quantization for speed-accuracy trade-off, demonstrates large zero-shot gains from low-FPS pretraining, and highlights challenges with mixed-precision fine-tuning for identity propagation.
174. ObjEmbed: Towards Universal Multimodal Object Embeddings
Core Problem: While existing multimodal embedding models excel at global image-text alignment, they struggle with fine-grained alignment between image regions and specific phrases, limiting their utility for object-oriented visual understanding tasks.
Key Innovation: ObjEmbed decomposes input images into multiple regional embeddings (for individual objects) and global embeddings. It captures semantic and spatial aspects via object and IoU embeddings, supports both region-level and image-level tasks, and efficiently encodes all objects in a single forward pass, demonstrating strong semantic discrimination.
175. Reg4Pru: Regularisation Through Random Token Routing for Token Pruning
Core Problem: Token pruning in Transformers, while improving computational efficiency, often leads to decreased stability of preserved representations and poorer dense prediction performance at deeper layers, especially for tasks like segmentation.
Key Innovation: Reg4Pru is a training regularisation technique that mitigates token-pruning performance loss for segmentation by using random token routing. It significantly improves average precision while achieving a relative speedup in wall-clock time, making token reduction strategies more viable for dense prediction.
176. SyNeT: Synthetic Negatives for Traversability Learning
Core Problem: Reliable traversability estimation for autonomous robots in complex outdoor environments is hindered by the lack of explicit negative data in existing self-supervised learning frameworks, limiting the model's ability to accurately identify diverse non-traversable regions.
Key Innovation: Introduces SyNeT, a method to explicitly construct synthetic negatives representing plausible but non-traversable regions, and integrates them into vision-based traversability learning to significantly enhance robustness and generalization across diverse environments.
177. TreeLoc: 6-DoF LiDAR Global Localization in Forests via Inter-Tree Geometric Matching
Core Problem: Reliable localization for navigation in forests is challenging due to degraded GPS and repetitive, occluded, and structurally complex LiDAR measurements, which weaken the assumptions of traditional urban-centric localization methods.
Key Innovation: Proposes TreeLoc, a LiDAR-based global localization framework for forests that handles place recognition and 6-DoF pose estimation by representing scenes using tree stems and their Diameter at Breast Height (DBH) for coarse and fine geometric matching.
178. Consideration of spatial variability and environmental impacts in the probabilistic design of driven piles
Core Problem: Traditional deterministic geotechnical designs for driven piles are overly conservative, leading to overdesign and increased environmental impact, by simplifying natural subsurface variability.
Key Innovation: A probabilistic framework for axial design of driven piles that incorporates subsurface spatial variability using Monte Carlo simulations, random field theory, and CPT-based methods, demonstrating material reductions and environmental savings while maintaining acceptable reliability.
179. A geometry consistent model for evaluating ship damaged stability at arbitrary attitudes based on the Quasi-Bonjean and the NSGA-II method
Core Problem: Conventional methods for evaluating ship damaged stability lack accuracy and adaptability under large heel/trim angles and complex flooding scenarios, which is critical for navigational safety in extreme marine conditions.
Key Innovation: A geometry-consistent model integrating a Quasi-Bonjean module with a multi-objective optimization (NSGA-II) framework is proposed, providing accurate and scalable evaluation of damaged stability at arbitrary attitudes, eliminating extrapolation errors, and supporting emergency decision-making.
180. Expert-guided and action-compensated deep reinforcement learning for robust multi-ship collision avoidance in dynamic and uncertain maritime environments
Core Problem: Ensuring safe navigation and effective collision avoidance for autonomous ships in complex, dynamic, and uncertain multi-ship maritime environments remains a fundamental challenge.
Key Innovation: An expert-guided and action-compensated deep reinforcement learning framework is proposed for robust multi-ship collision avoidance, integrating expert trajectories for global guidance, a line-of-sight/PID action compensation mechanism for smooth control, and an enhanced collision risk modeling approach, significantly outperforming conventional methods.
181. From single-modal to multi-modal: How does multi-modal data integration enhance the precision of seafarer fatigue detection?
Core Problem: Seafarer fatigue significantly increases accident risk, but existing detection methods lack accuracy and reliability due to reliance on single-modal, simulated data.
Key Innovation: A multi-modal data integration approach (EEG, EDA, ECG, Psych) from a real navigation experiment, combined with a feature layer fusion strategy and machine learning (LightGBM), achieved 95.93% accuracy in seafarer fatigue detection, significantly outperforming single-modal methods and enhancing navigation safety.
182. Protective performance of biomimetic foam buffers in RC piers against ship collisions
Core Problem: RC piers in inland waterways are vulnerable to damage from ship collisions, necessitating effective protection devices to enhance safety and structural integrity.
Key Innovation: A modular biomimetic foam buffer device (ACF-filled square tubes) is proposed, demonstrating effective energy absorption (reducing peak impact force by 60.38%) and improved lateral stiffness retention, shifting pier failure mode from shear to bending, thus providing an effective solution for bridge collision protection.
183. Assessment of human contribution to very serious maritime accidents based on machine learning techniques
Core Problem: Quantifying human contributions to very serious maritime accidents is crucial for maritime safety management, but existing methods may lack systematic identification and accurate assessment of human factors.
Key Innovation: Machine learning techniques (LightGBM with Association Rule) are used to quantify human contributions to very serious maritime accidents, achieving 85.94% accuracy and identifying key human factors (e.g., failure to follow rules, inadequate safety management) for prioritizing safety interventions.
184. SEEPS4ALL: an open dataset for the verification of daily precipitation forecasts using station climate statistics
Core Problem: The computation of precipitation forecast verification scores like SEEPS is not straightforward due to the requirement for detailed precipitation climatology at verification locations, limiting their widespread use.
Key Innovation: Introduction of SEEPS4ALL, an open dataset and set of tools that democratize the use of climate statistics for verifying daily precipitation forecasts, showcasing its application for both deterministic and probabilistic forecasts.
185. Joint calibration of multi-scale hydrological data sets using probabilistic water balance data fusion: methodology and application to the irrigated Hindon River Basin, India
Core Problem: Hydrological datasets are subject to uncertainties, limiting their potential for water resource management.
Key Innovation: Developed and applied a monthly probabilistic water balance data fusion approach for automatic bias correction and noise filtering of multi-scale hydrological data, generating hydrologically consistent estimates of water balance components and reducing uncertainties.
186. Technical note: Including hydrologic impact definition in climate projection uncertainty partitioning: a case study of the Central American mid-summer drought
Core Problem: Understanding and quantifying uncertainty in projected hydrological impacts (specifically mid-summer drought) under climate change, given sensitivity to MSD definition.
Key Innovation: Characterized contributions to total uncertainty from MSD definition, downscaling method, and climate models in climate projection uncertainty partitioning for the Central American mid-summer drought, providing guidance for water planning and adaptation efforts.
187. Questioning the Endorheic Paradigm: water balance dynamics in the Salar del Huasco basin, Chile
Core Problem: Understanding how rainfall and evaporation drive spatial and temporal dynamics of groundwater recharge and water balance in arid endorheic basins, and potentially challenging the endorheic assumption.
Key Innovation: Implemented a modified semi-distributed rainfall-runoff model with satellite data to examine water balance dynamics in an arid endorheic basin, revealing spatial patterns of groundwater recharge and evaporation, and challenging the basin's endorheic assumption due to unaccounted water loss/inflow.
188. Integrating SDGSAT-1 glimmer imagery with Sentinel-1/2 data for high-resolution building height estimation
Core Problem: Accurately mapping building heights over wide areas at fine scales remains challenging due to high costs, loss of spatial details in large-scale mapping, and underestimation of high-rise buildings.
Key Innovation: A novel Tri-Modal Height Estimation Network (TMHEN) and Cross-modal Adaptive Enhancer (CAE) module integrating SDGSAT-1 Glimmer imagery (first time used), Sentinel-2 MSI, and Sentinel-1 SAR data to derive high-resolution (10-m) building height maps, improving accuracy and mitigating underestimation.
189. Sea ice thickness surveys with a drone-borne multi-frequency EM sensor
Core Problem: Climate change-induced variations in sea ice conditions make traditional sled-towed EM surveys hazardous, necessitating safer and more efficient methods for sea ice thickness monitoring.
Key Innovation: Developed and tested a drone-borne multi-frequency EM sensor combined with custom altitude/attitude sensors to remotely measure sea ice thickness, successfully retrieving data in agreement with manual measurements and demonstrating strong potential for efficient and accurate surveys from a safe distance.
190. Linking Gondwana inheritance to Alpine paleogeography in the Northern Dora-Maira Massif (Western Alps)
Core Problem: The influence of inherited structures in rifted continental margins on the architecture and evolution of collisional orogens, specifically in the Western Alps, needs further clarification.
Key Innovation: New lithostratigraphic, structural data, and U–Pb zircon dating in the Dora-Maira Massif revealed a long-lasting tectonostratigraphic/magmatic evolution from pre-Permian to Jurassic. The study highlights that pre-rift architecture governed margin segmentation, and successive cycles created structural and rheological heterogeneities that likely localized strain during Cenozoic Alpine overprinting, offering a broader framework for understanding orogen dynamics.
191. Potassium isotope characteristics of typical salt lake in key tectonic zones, China: Sources and evolutionary models
Core Problem: The sources and evolutionary models of potassium in salt lakes situated within critical tectonic zones, particularly regarding the role of deep geological processes, are not fully understood.
Key Innovation: This study investigated potassium isotopic compositions (δ41K) in Lakkor Co Salt Lake (Qinghai-Tibet Plateau). Results showed distinct K isotope fractionation, with recharge rivers having lower δ41K values than global averages, and surface brine falling within modern seawater/Qaidam Basin ranges. Combined hydrochemical, δ11B, and δ7Li data indicated K sources from surface rock weathering, geothermal fluid, and potentially ultra-high pressure metamorphic zones formed by deep subduction of oceanic crust, providing a theoretical basis for analogous salt lake exploration and understanding elemental behavior in subducting oceanic crust.
192. Robust discharge prediction of seasonal snow-influenced karst systems through hybridization of process-based and data-driven models
Core Problem: Hydrological modeling of karst systems, especially those influenced by seasonal snow cover, is challenging due to their dynamic, nonlinear behavior and the difficulty in predicting discharge during extreme flow conditions.
Key Innovation: Development of an innovative hybrid modeling approach combining a process-based model (CemaNeige GR6J) and a data-driven model (SAE-DNN), which significantly outperforms stand-alone models in robustly predicting daily discharge, particularly during extreme flow conditions, for snow-influenced karst systems.
193. Integrated multi-scale ecohydrogeological monitoring of spatio-temporal dynamics in karst critical zones
Core Problem: The vulnerability of karst environments to changing hydrometeorological patterns and vegetation disturbance necessitates a unified, interdisciplinary strategy for comprehensive understanding and monitoring of spatio-temporal dynamics in karst critical zones (KCZ).
Key Innovation: An integrated multi-scale ecohydrogeological monitoring approach combining surface and underground sites with advanced methods (enhanced precipitation monitoring, customized lysimeter techniques, microscale cave adaptations) to decipher flow dynamics and improve data collection representativeness in heterogeneous karst environments.
194. Comparative analysis of GAMLSS modeling approaches for nonstationary runoff dynamics in the Yellow River Basin of China
Core Problem: Systematically assessing nonstationary characteristics and multiple driving mechanisms (climate change, human activities) of runoff dynamics at the basin scale, and improving runoff modeling and drought/flood detection accuracy.
Key Innovation: Comparison of two GAMLSS modeling approaches (Continuous-series and monthly-segmented) for nonstationary runoff dynamics in the Yellow River Basin, demonstrating that monthly-segmented modeling (Mode2) significantly enhances model robustness, better captures seasonal dynamics and human influences, and is more suitable for refined water resource management and extreme drought-flood prediction.
195. Simulated Changes and Future Analogy Extent of Ocean Heat Content During the Mid‐Pliocene Warm Period
Core Problem: Understanding how globally warmer climates store oceanic heat is crucial for future climate projections, and the Mid-Pliocene Warm Period offers an opportunity for this analogy.
Key Innovation: Uses the PlioMIP2 model ensemble to quantify global ocean heat content (OHC) during the Mid-Pliocene Warm Period, finding it globally higher than pre-industrial and exceeding highest SSP5-8.5 future scenarios, implying substantial oceanic heat absorption capacity.
196. “Hearing” Wind Speed: Ground Wind Measurement Using Deep Learning From Surveillance Audio
Core Problem: There is a need for high-resolution, low-cost urban ground wind observations for various applications, including urban weather forecasting and air pollution assessment.
Key Innovation: Develops a novel deep learning method using surveillance audio and continuous wavelet transform to measure ground wind speed with high accuracy (84.56% prediction accuracy, RMSE of 1.84 m/s for typhoons), demonstrating the potential for a cost-effective urban wind observation network.
197. Recent Cloud Controlling Factor Analyses Indicate Higher Climate Sensitivity
Core Problem: Cloud feedback is a dominant source of uncertainty in climate model estimates of equilibrium climate sensitivity (ECS), making future climate projections uncertain.
Key Innovation: Uses separate frameworks for high- and low-cloud feedbacks in cloud controlling factor analysis, yielding robustly positive estimates of overall cloud feedback, which, when applied as constraints, indicate a shift toward higher equilibrium climate sensitivity with reduced uncertainty.
198. Origins of Precipitation in the World's Water Towers
Core Problem: Poorly quantified moisture sources for precipitation in high-mountain 'water towers,' which are critical for freshwater supply and are exacerbated by climate change.
Key Innovation: Combined two atmospheric moisture tracking methods to reveal that terrestrial evaporation contributes approximately half of total precipitation in water towers, with inland WTUs relying more on land-sourced moisture, and oceanic moisture driving snowfall while terrestrial transpiration sustains rainfall.
199. Evaporation‐Induced Hysteresis in Surface Water‐Groundwater Exchange in Wetlands
Core Problem: Poorly understood influence of evaporation on subsurface feedbacks and surface water-groundwater exchange in wetlands.
Key Innovation: Used an integrated surface-subsurface hydrologic and solute transport model to show that evaporation can induce hysteresis in groundwater/solute upwelling, with strength governed by sediment permeability and site-specific conditions, highlighting evaporation as a critical indirect driver of wetland water/solute exchange.
200. A Semi-Supervised Pipeline for Generalized Behavior Discovery from Animal-Borne Motion Time Series
Core Problem: Learning behavioral taxonomies from animal-borne motion time series is challenging due to scarce labels, class imbalance, and the presence of novel, unannotated behaviors.
Key Innovation: Proposes a semi-supervised pipeline for generalized behavior discovery that learns embeddings, performs label-guided clustering, and introduces a KDE + HDR (highest-density region) containment score to quantitatively detect and flag truly novel behaviors in ecological motion time series.
201. SVD-ViT: Does SVD Make Vision Transformers Attend More to the Foreground?
Core Problem: Vision Transformers (ViT) lack an explicit mechanism to distinguish foreground from background, leading to learning unnecessary background features and degraded classification performance.
Key Innovation: SVD-ViT, which leverages singular value decomposition (SVD) with SPC module, SSVA, and ID-RSVD components to prioritize learning foreground features, improving classification accuracy and reducing background noise impact.
202. Super-Resolution and Denoising of Corneal B-Scan OCT Imaging Using Diffusion Model Plug-and-Play Priors
Core Problem: High-speed OCT acquisitions degrade spatial resolution and increase speckle noise, challenging accurate interpretation of corneal images.
Key Innovation: An advanced super-resolution framework leveraging diffusion model plug-and-play (PnP) priors to achieve 4x spatial resolution enhancement and effective denoising of OCT B-scan images, formulated as a Bayesian inverse problem.
203. SC3D: Dynamic and Differentiable Causal Discovery for Temporal and Instantaneous Graphs
Core Problem: Discovering causal structures from multivariate time series is challenging due to interactions across multiple lags and instantaneous dependencies, and the combinatorial search space of dynamic graphs.
Key Innovation: Proposes SC3D, a two-stage differentiable framework that jointly learns lag-specific adjacency matrices and instantaneous directed acyclic graphs, achieving improved stability and more accurate recovery of causal structures.
204. Semantics-Aware Generative Latent Data Augmentation for Learning in Low-Resource Domains
Core Problem: Deep learning underperforms in data-scarce settings, and foundation models still suffer from scarce labeled data during downstream fine-tuning.
Key Innovation: Proposes GeLDA, a semantics-aware generative latent data augmentation framework that leverages conditional diffusion models to synthesize samples in an FM-induced latent space, improving performance in zero-shot speech emotion recognition and long-tailed image classification.
205. Weighted Temporal Decay Loss for Learning Wearable PPG Data with Sparse Clinical Labels
Core Problem: Developing health algorithms from wearable PPG biosignals is challenged by the sparsity of clinical labels, making temporally distant biosignals less reliable for supervision.
Key Innovation: Introduces a simple training strategy that learns a biomarker-specific decay of sample weight over the time gap between a segment and its ground truth label, improving AUPRC over baselines and providing interpretable temporal sensitivity.
206. A Reproducible Framework for Bias-Resistant Machine Learning on Small-Sample Neuroimaging Data
Core Problem: Conventional cross-validation frameworks for small-sample neuroimaging data yield optimistically biased results, limiting reproducibility and generalization.
Key Innovation: Introduces a reproducible, bias-resistant ML framework integrating domain-informed feature engineering, nested cross-validation, and calibrated decision-threshold optimization, achieving unbiased evaluation and interpretability on high-dimensional neuroimaging data.
207. Dynamic High-frequency Convolution for Infrared Small Target Detection
Core Problem: Single-frame infrared small target (SIRST) detection is challenging because many high-frequency components (HFCs) like clutter resemble targets, and current learning methods neglect explicit modeling and discriminative representation of these HFCs.
Key Innovation: Proposes Dynamic High-frequency Convolution (DHiF), which translates discriminative modeling into generating a dynamic local filter bank. DHiF is sensitive to HFCs due to dynamically adjusted parameters, adaptively processing different HFC regions to capture distinctive grayscale variations, leading to superior detection performance.
208. Fisheye Stereo Vision: Depth and Range Error
Core Problem: Understanding and quantifying the depth and range error in fisheye stereo vision systems, especially at large angles, is crucial for accurate 3D reconstruction.
Key Innovation: Derives analytical expressions for the depth and range error of fisheye stereo vision systems as a function of object distance, specifically accounting for accuracy at large angles.
209. From Zero to Hero: Advancing Zero-Shot Foundation Models for Tabular Outlier Detection
Core Problem: Effective deployment of outlier detection (OD) on new tasks is hindered by the lack of labeled outliers, making algorithm and hyperparameter selection difficult, and existing foundation models for OD can be further advanced.
Key Innovation: Introduces OUTFORMER, an advancement in zero-shot foundation models for tabular outlier detection. It uses a mixture of synthetic priors and self-evolving curriculum training, achieving state-of-the-art performance on multiple benchmarks with fast, zero-shot inference, requiring no labeled outliers for new tasks.
210. SAFE-KD: Risk-Controlled Early-Exit Distillation for Vision Backbones
Core Problem: Practical deployment of early-exit networks for reducing inference cost is hindered by the lack of reliable mechanisms to determine when an early exit is safe, risking misclassification.
Key Innovation: SAFE-KD, a universal multi-exit wrapper that couples hierarchical distillation with conformal risk control to calibrate per-exit stopping thresholds, guaranteeing a user-specified selective misclassification risk and yielding improved accuracy-compute trade-offs and robust performance.
211. IVC-Prune: Revealing the Implicit Visual Coordinates in LVLMs for Vision Token Pruning
Core Problem: Large Vision-Language Models (LVLMs) have prohibitive inference costs for high-resolution visual inputs, and existing visual token pruning methods often discard tokens crucial for spatial reasoning by focusing solely on semantic relevance.
Key Innovation: IVC-Prune, a training-free, prompt-aware pruning strategy that identifies and retains 'implicit visual coordinates' (IVC tokens) derived from Rotary Position Embeddings (RoPE) and semantically relevant foreground tokens, reducing visual tokens by approximately 50% while maintaining or improving performance and spatial reasoning.
212. TextME: Bridging Unseen Modalities Through Text Descriptions
Core Problem: Expanding multimodal representations to novel modalities is constrained by the high cost and infeasibility of acquiring large-scale paired datasets for new modality combinations, especially in specialized domains.
Key Innovation: TextME, the first text-only modality expansion framework, which projects diverse modalities into LLM embedding space as a unified anchor by exploiting the geometric structure of pretrained contrastive encoders, enabling zero-shot cross-modal transfer and emergent cross-modal retrieval without paired supervision.
213. SwiftVLM: Efficient Vision-Language Model Inference via Cross-Layer Token Bypass
Core Problem: Existing visual token pruning methods for Vision-Language Models (VLMs) often make early pruning decisions that lead to irreversible loss of critical information, causing performance degradation on tasks requiring fine-grained visual details.
Key Innovation: SwiftVLM, a training-free method that introduces a 'bypass' pruning paradigm, preserving unselected visual tokens and forwarding them for re-evaluation in subsequent layers, enabling independent pruning decisions across layers and achieving superior accuracy-efficiency trade-offs for VLMs.
214. SATORIS-N: Spectral Analysis based Traffic Observation Recovery via Informed Subspaces and Nuclear-norm minimization
Core Problem: Accurately imputing partially observed traffic-density matrices, especially under high occlusion, is challenging for critical applications in intelligent vehicle systems.
Key Innovation: SATORIS-N, a framework that uses a subspace-aware semidefinite programming formulation of nuclear norm to explicitly inform reconstruction with prior singular-subspace information, achieving robust and accurate imputation of spatiotemporal traffic data.
215. LaVPR: Benchmarking Language and Vision for Place Recognition
Core Problem: Visual Place Recognition (VPR) often fails under extreme environmental changes and perceptual aliasing, and standard systems cannot perform 'blind' localization from verbal descriptions, a capability needed for applications such as emergency response.
Key Innovation: Introduces LaVPR, a large-scale benchmark extending VPR datasets with natural-language descriptions, enabling Multi-Modal Fusion for enhanced robustness and Cross-Modal Retrieval for language-based localization, yielding consistent gains in visually degraded conditions.
216. PWAVEP: Purifying Imperceptible Adversarial Perturbations in 3D Point Clouds via Spectral Graph Wavelets
Core Problem: Recent progress in adversarial attacks on 3D point clouds, particularly in achieving spatial imperceptibility and high attack performance, presents significant challenges for defenders, as current defensive approaches are cumbersome and invasive.
Key Innovation: Proposes PWAVEP, a plug-and-play and non-invasive defense mechanism in the spectral domain, which purifies imperceptible adversarial perturbations by computing spectral graph wavelet domain saliency and local sparsity scores, then hierarchically eliminating salient points and applying spectral filtering to attenuate high-frequency adversarial noise.
217. Tiled Prompts: Overcoming Prompt Underspecification in Image and Video Super-Resolution
Core Problem: Text-conditioned diffusion models for image and video super-resolution, when scaled to high resolutions using latent tiling, suffer from prompt underspecification where a single global caption misses localized details and provides locally irrelevant guidance.
Key Innovation: Proposes Tiled Prompts, a unified framework that generates a tile-specific prompt for each latent tile and performs super-resolution under locally text-conditioned posteriors, providing high-information guidance that resolves prompt underspecification with minimal overhead, leading to consistent gains in perceptual quality and text alignment.
218. Socratic-Geo: Synthetic Data Generation and Geometric Reasoning via Multi-Agent Interaction
Core Problem: Multimodal Large Language Models (MLLMs) struggle with geometric reasoning due to the extreme scarcity of high-quality image-text pairs, and existing automated data generation methods fail to ensure fidelity and training effectiveness.
Key Innovation: Proposes Socratic-Geo, a fully autonomous framework that dynamically couples data synthesis with model learning through multi-agent interaction. A Teacher agent generates parameterized Python scripts with reflective feedback, and a Solver agent optimizes reasoning through preference learning, achieving state-of-the-art in geometric reasoning and image generation.
219. Inlier-Centric Post-Training Quantization for Object Detection Models
Core Problem: Object detection models have immense computational demands, and quantization is complicated by task-irrelevant activations (anomalies) that expand activation ranges and skew distributions, making bit allocation difficult.
Key Innovation: Presents InlierQ, an inlier-centric post-training quantization approach that separates anomalies from informative inliers using gradient-aware volume saliency scores and a posterior distribution fitting. This suppresses anomalies while preserving informative features, consistently reducing quantization error.
220. DeepDFA: Injecting Temporal Logic in Deep Learning for Sequential Subsymbolic Applications
Core Problem: Integrating logical knowledge into deep neural network training is a hard challenge, especially for sequential or temporally extended domains involving subsymbolic observations.
Key Innovation: Proposes DeepDFA, a neurosymbolic framework that integrates high-level temporal logic (Deterministic Finite Automata or Moore Machines) into neural architectures. It models temporal rules as continuous, differentiable layers, enabling symbolic knowledge injection and outperforming traditional deep learning models in temporal knowledge integration.
221. CoGenCast: A Coupled Autoregressive-Flow Generative Framework for Time Series Forecasting
Core Problem: Time series forecasting requires both semantic understanding over contextual conditions and stochastic modeling of continuous temporal dynamics, but existing approaches (LLMs or diffusion models) struggle to adequately model both simultaneously.
Key Innovation: Proposes CoGenCast, a hybrid generative framework that couples pre-trained LLMs (reconfigured for forecasting) with a flow-matching mechanism to effectively model both semantic context and continuous stochastic temporal dynamics, achieving state-of-the-art performance in time series forecasting.
222. Multi-Objective Optimization for Synthetic-to-Real Style Transfer
Core Problem: Semantic segmentation networks trained on synthetic data perform poorly on real images due to the domain gap, and choosing effective style transfer pipelines to bridge this gap is challenging due to a large combinatorial search space.
Key Innovation: Formulation of style transfer as a sequencing problem for multi-objective evolutionary optimization, using genetic algorithms to balance structural coherence and style similarity, and studying efficient paired-image metrics for rapid pipeline evaluation in synthetic-to-real domain adaptation.
223. Data-Driven Graph Filters via Adaptive Spectral Shaping
Core Problem: The need for data-driven graph filtering methods that are scalable, interpretable, and can generalize across different graphs, while effectively allocating energy to heterogeneous regions of the Laplacian spectrum.
Key Innovation: Introduction of Adaptive Spectral Shaping, a data-driven framework for graph filtering that learns a reusable baseline spectral kernel modulated by Gaussian factors, and Transferable Adaptive Spectral Shaping (TASS) for few-shot transfer, combining scalability, interpretability, and cross-graph generalization.
224. RAWDet-7: A Multi-Scenario Benchmark for Object Detection and Description on Quantized RAW Images
Core Problem: Most vision models discard valuable sensor-level information by training on processed RGB images, hindering fine-grained object detection and description, especially under realistic sensor constraints like low-bit quantization.
Key Innovation: Introduction of RAWDet-7, a large-scale multi-scenario benchmark dataset of RAW images with dense annotations and object-level descriptions, enabling research into object detection, description quality, and generalization using unprocessed and quantized RAW image data.
225. FOVI: A biologically-inspired foveated interface for deep vision models
Core Problem: Computer vision systems struggle to efficiently process full-field high-resolution images due to uniform resolution encoding, unlike human foveated vision.
Key Innovation: FOVI, a biologically-inspired foveated vision interface, reformats variable-resolution sensor data into a uniformly dense manifold, enabling efficient kNN-convolution and foveated adaptation of models like DINOv3 ViT, achieving competitive performance with reduced computational cost.
226. Reward Redistribution for CVaR MDPs using a Bellman Operator on L-infinity
Core Problem: Static Conditional Value-at-Risk (CVaR) objectives in safety-critical applications lack a recursive Bellman decomposition in MDPs, and existing state augmentation methods lead to sparse rewards and degenerate fixed points.
Key Innovation: A novel formulation of the static CVaR objective based on augmentation, leading to a Bellman operator with dense per-step rewards and contracting properties, enabling the development of risk-averse value iteration and Q-learning algorithms with convergence guarantees.
227. Enhancing Imbalanced Node Classification via Curriculum-Guided Feature Learning and Three-Stage Attention Network
Core Problem: Imbalanced node classification in Graph Neural Networks (GNNs) leads to unfair learning and poor performance on minority classes.
Key Innovation: CL3AN-GNN, a Curriculum-Guided Feature Learning and Three-Stage Attention Network, uses a three-step attention system (Engage, Enact, Embed) to progressively learn from simpler to more complex features, effectively addressing label skew and improving performance on imbalanced node classification tasks.
228. SymPlex: A Structure-Aware Transformer for Symbolic PDE Solving
Core Problem: Discovering analytical symbolic solutions to Partial Differential Equations (PDEs) without ground-truth expressions is challenging, and existing numerical/neural methods approximate solutions rather than providing interpretable symbolic forms.
Key Innovation: SymPlex, a reinforcement learning framework with a structure-aware Transformer (SymFormer), formulates symbolic PDE solving as tree-structured decision-making, optimizing candidate solutions using only PDE and boundary conditions to discover interpretable, human-readable analytical symbolic solutions.
229. EventNeuS: 3D Mesh Reconstruction from a Single Event Camera
Core Problem: Existing event-based techniques are severely limited in their 3D reconstruction accuracy from event cameras.
Key Innovation: EventNeuS, a self-supervised neural model that combines 3D signed distance function and density field learning with event-based supervision and spherical harmonics encodings, significantly improving 3D mesh reconstruction accuracy from monocular color event streams.
230. From Task Solving to Robust Real-World Adaptation in LLM Agents
Core Problem: Existing evaluations for LLM agents often assume 'clean interfaces' and stable environments, overestimating real-world readiness, where agents face underspecified rules, unreliable signals, shifting environments, and implicit goals.
Key Innovation: A stress-testing framework for deployment-relevant robustness of LLM agents under partial observability, dynamic environments, noisy signals, and dynamic agent state, using a grid-based game to reveal large gaps between nominal task-solving and robust adaptation, and identifying model-specific sensitivities and failure drivers.
231. Training-Free Self-Correction for Multimodal Masked Diffusion Models
Core Problem: Masked diffusion models suffer from error accumulation during sampling because they update multiple tokens simultaneously and treat generated tokens as immutable, making early mistakes uncorrectable.
Key Innovation: Proposes a training-free self-correction framework that exploits the inductive biases of pre-trained masked diffusion models, significantly improving generation quality on text-to-image and multimodal understanding tasks with reduced sampling steps, without modifying model parameters or requiring auxiliary evaluators.
232. Weighted Sum-of-Trees Model for Clustered Data
Core Problem: Existing methods for clustered data (e.g., mixed models, extensions to trees) are limited in predicting for out-of-sample groups, often assuming a common outcome model across clusters, and struggle to infer similarity across groups in the outcome prediction model.
Key Innovation: Proposes a weighted sum-of-trees model that learns a distinct decision tree for each sample group and combines their predictions using weights based on group similarity, outperforming traditional methods in simulations and allowing for inference on inter-group similarity.
233. Online Conformal Prediction via Universal Portfolio Algorithms
Core Problem: Existing Online Conformal Prediction (OCP) methods, which aim for long-run coverage for arbitrary data streams, often require manual learning-rate tuning and lack a general theoretical framework for coverage bounds.
Key Innovation: Develops a general regret-to-coverage theory for interval-valued OCP based on the (1-alpha)-pinball loss, identifying linearized regret as key, and proposes UP-OCP, a parameter-free method via reduction to a two-asset portfolio selection problem, demonstrating strong finite-time bounds and superior size/coverage trade-offs.
234. NeuralFLoC: Neural Flow-Based Joint Registration and Clustering of Functional Data
Core Problem: Clustering functional data is challenging due to phase variation (temporal misalignment) obscuring intrinsic shape differences, and most existing approaches treat registration and clustering separately or rely on restrictive parametric assumptions.
Key Innovation: Presents NeuralFLoC, a fully unsupervised, end-to-end deep learning framework for joint functional registration and clustering based on Neural ODE-driven diffeomorphic flows and spectral clustering, which learns smooth warping functions and cluster-specific templates simultaneously, achieving state-of-the-art performance and robustness.
235. Principled Federated Random Forests for Heterogeneous Data
Core Problem: Existing federated learning methods for Random Forests rely on unprincipled heuristics and fail to optimize the global impurity criterion, especially under distribution shifts and diverse client data heterogeneity.
Key Innovation: FedForest, a new federated RF algorithm for horizontally partitioned data, approximates the centralized split selection by aggregating client statistics, naturally accommodates diverse data heterogeneity, and allows for non-parametric personalization, closely matching centralized performance.
236. PACE: Pretrained Audio Continual Learning
Core Problem: Pretrained audio models are fragile in real-world settings where data distributions shift over time, and existing continual learning (CL) strategies from vision lead to poor performance in audio due to upstream-downstream misalignment.
Key Innovation: PACE, a novel method that enhances analytic classifiers with first-session adaptation (FSA) via regularization, enables multi-session adaptation through adaptive subspace-orthogonal PEFT, and introduces spectrogram-based boundary-aware perturbations, substantially outperforming state-of-the-art baselines in audio CL.
237. Improving the Linearized Laplace Approximation via Quadratic Approximations
Core Problem: The Linearized Laplace Approximation (LLA) for Bayesian uncertainty quantification in Deep Neural Networks (DNNs) uses a linearized model for prediction, which can degrade fidelity to the true Laplace approximation.
Key Innovation: The Quadratic Laplace Approximation (QLA), which approximates each second-order factor in the approximate Laplace log-posterior using a rank-one factor, yielding modest yet consistent uncertainty estimation improvements over LLA without significantly increasing computational cost.
238. Generator-based Graph Generation via Heat Diffusion
Core Problem: Existing graph generative models may lack a unified framework that injects domain-specific inductive bias while retaining neural approximator flexibility, or struggle to effectively capture structural properties.
Key Innovation: A novel framework for graph generation adapting Generator Matching to graph-structured data, leveraging graph Laplacian and heat kernel to define continuous-time diffusion, unifying and generalizing existing diffusion-based models.
239. Improved Analysis of the Accelerated Noisy Power Method with Applications to Decentralized PCA
Core Problem: Previous analyses of the Accelerated Noisy Power Method for Principal Component Analysis (PCA) required overly restrictive conditions on perturbations, limiting its practical applicability, especially in decentralized settings.
Key Innovation: An improved analysis of the Accelerated Noisy Power Method, preserving accelerated convergence under much milder noise conditions, and deriving the first provably accelerated algorithm for decentralized PCA with similar communication costs.
240. Preference-based Conditional Treatment Effects and Policy Learning
Core Problem: Existing methods for conditional treatment effect estimation and policy learning struggle with multivariate, ordinal, or preference-driven outcomes, and face intrinsic non-identifiability for comparison-based estimands.
Key Innovation: A new preference-based framework for conditional treatment effect estimation and policy learning, built on Conditional Preference-based Treatment Effect (CPTE), providing interpretable targets, new identifiability conditions, and efficient influence-function estimators.
241. Discrete Latent Structure in Neural Networks
Core Problem: Learning with discrete latent structures in neural networks is challenging because neural networks are typically designed for continuous computation, making effective training difficult.
Key Innovation: A comprehensive exploration of three broad strategies for learning with discrete latent structure (continuous relaxation, surrogate gradients, probabilistic estimation), revealing new connections and providing a consistent notation for a wide range of models.
242. Sparse maximal update parameterization: A holistic approach to sparse training dynamics
Core Problem: Sparse neural networks face challenges including impaired signal propagation, high tuning costs due to new hyperparameters, and suboptimal performance compared to dense models because optimal hyperparameters differ.
Key Innovation: Proposes SμPar, a holistic reparameterization approach for random unstructured static sparsity, ensuring activations, gradients, and weight updates scale independently of sparsity level and allowing hyperparameters tuned on small dense networks to transfer to large sparse models, significantly reducing tuning costs and improving performance.
243. Conformal Prediction for Causal Effects of Continuous Treatments
Core Problem: Existing conformal prediction methods for causal effects are limited to binary/discrete treatments and require known propensity scores, making them unsuitable for continuous treatments or scenarios where propensity scores must be estimated.
Key Innovation: Proposes a novel conformal prediction method for potential outcomes of continuous treatments, which accounts for the uncertainty introduced by propensity estimation, providing valid finite-sample prediction intervals even when propensity scores are unknown.
244. Fast Training of Sinusoidal Neural Fields via Scaling Initialization
Core Problem: Sinusoidal Neural Fields (SNFs) suffer from high training costs, and their standard initialization scheme is suboptimal, hindering broader adoption.
Key Innovation: Proposes "weight scaling" (multiplying each weight, except the last layer, by a constant) to accelerate SNF training by 10x, consistently providing significant speedup and outperforming more recently proposed architectures by resolving spectral bias and ensuring a well-conditioned optimization trajectory.
245. Deep Graph Learning will stall without Network Science
Core Problem: Deep graph learning prioritizes empirical performance and overlooks fundamental insights from network science, potentially leading to a stagnation in its progress despite sharing the goal of modeling graph-structured data.
Key Innovation: A position paper arguing for the necessity of integrating insights from network science into deep graph learning to address current issues and ensure continued progress, proposing six "Calls for Action" to bridge this gap.
246. FedVSR: Towards Model-Agnostic Federated Learning in Video Super-Resolution
Core Problem: Federated learning (FL) struggles with low-level vision tasks like video super-resolution (VSR), producing blurry outputs, and existing FL frameworks are not specifically designed for VSR, despite the privacy benefits of decentralized data.
Key Innovation: Introduces FedVSR, the first model-agnostic and stateless federated learning framework for VSR, which employs a lightweight Discrete Wavelet Transform (DWT)-based loss function for high-frequency detail preservation and a loss-aware aggregation strategy, significantly improving perceptual video quality with minimal computational and communication overhead.
247. Seeing through Satellite Images at Street Views
Core Problem: Rendering photorealistic street-view panoramas and videos from satellite images is challenging due to the sparse-view nature and extremely large viewpoint changes between satellite and street-view images.
Key Innovation: Proposes Sat2Density++, a novel approach that models street-view specific elements (sky, illumination) within neural radiance fields, enabling the rendering of photorealistic, multi-view consistent street-view panoramas faithful to the satellite image.
248. Multi-view Graph Condensation via Tensor Decomposition
Core Problem: Training Graph Neural Networks (GNNs) on large-scale graphs presents significant computational challenges, and existing graph condensation methods are computationally intensive, lack interpretability, and do not maintain a mapping between synthetic and original nodes.
Key Innovation: Proposes Multi-view Graph Condensation via Tensor Decomposition (GCTD) to synthesize smaller, informative graphs, reducing computational demands while preserving GNN performance and offering a more transparent alternative to bi-level optimization.
249. DiffVL: Diffusion-Based Visual Localization on 2D Maps via BEV-Conditioned GPS Denoising
Core Problem: Accurate visual localization for autonomous driving is challenged by the high cost of HD maps and the unreliability of noisy GPS signals in urban environments, which existing SD-map-based methods often overlook.
Key Innovation: Proposes DiffVL, a diffusion-based framework that reformulates visual localization as a GPS denoising task, recovering true pose distribution by iteratively refining noisy GPS trajectories conditioned on visual BEV features and SD maps, achieving sub-meter accuracy without HD maps.
250. Exploring the Global-to-Local Attention Scheme in Graph Transformers: An Empirical Study
Core Problem: Existing Graph Transformers (GTs) with local-and-global or local-to-global attention schemes may suffer from information loss, where local neighborhood information from GNNs is diluted by global attention mechanisms primarily capturing long-range dependencies.
Key Innovation: Proposes G2LFormer, a novel global-to-local attention scheme where shallow layers capture global information via attention and deeper layers learn local structural information via GNNs, preventing dilution of local neighborhood information, and introduces a cross-layer information fusion strategy.
251. Transformers can do Bayesian Clustering
Core Problem: Bayesian clustering is computationally demanding at scale, and real-world datasets often contain missing values, leading to suboptimal results when uncertainty is ignored.
Key Innovation: Introduces Cluster-PFN, a Transformer-based model that extends Prior-Data Fitted Networks (PFNs) for unsupervised Bayesian clustering, accurately estimating the posterior distribution over the number of clusters and assignments, outperforming handcrafted methods and imputation-based baselines while being orders of magnitude faster.
252. Dynamic Priors in Bayesian Optimization for Hyperparameter Optimization
Core Problem: Most Bayesian Optimization (BO) methods for hyperparameter optimization (HPO) only incorporate expert knowledge during initialization, limiting user influence as new insights emerge in iterative machine learning development.
Key Innovation: DynaBO, a BO framework, enables continuous user control by augmenting the acquisition function with decaying, prior-weighted preferences, while preserving asymptotic convergence and introducing a data-driven safeguard against misleading priors, outperforming state-of-the-art competitors.
253. DiScoFormer: Plug-In Density and Score Estimation with Transformers
Core Problem: Existing methods for estimating probability density and its score from samples are bifurcated: classical kernel density estimators (KDE) suffer from the curse of dimensionality, while neural score models require retraining for every target distribution.
Key Innovation: DiScoFormer, an equivariant Transformer, provides a "train-once, infer-anywhere" solution that maps i.i.d. samples to both density values and score vectors, generalizing across distributions and sample sizes, converging faster and achieving higher precision than KDE.
254. Polynomial Neural Sheaf Diffusion: A Spectral Filtering Approach on Cellular Sheaves
Core Problem: Common Neural Sheaf Diffusion implementations for Graph Neural Networks suffer from limitations like SVD-based normalization, dense per-edge restriction maps (scaling with stalk dimension), frequent Laplacian rebuilds, and brittle gradients, hindering scalability and stability.
Key Innovation: PolyNSD, a new sheaf diffusion approach, uses a degree-K polynomial in a normalized sheaf Laplacian, evaluated via a stable three-term recurrence. This provides an explicit K-hop receptive field, trainable spectral response, and enforces stability, achieving new state-of-the-art results on benchmarks with reduced runtime and memory.
255. MSACL: Multi-Step Actor-Critic Learning with Lyapunov Certificates for Exponentially Stabilizing Control
Core Problem: Model-free reinforcement learning (RL) for safety-critical applications struggles with establishing verifiable stability guarantees while maintaining high exploration efficiency, often relying on complex reward engineering and single-step constraints.
Key Innovation: MSACL (Multi-Step Actor-Critic Learning with Lyapunov Certificates) integrates exponential stability with maximum entropy reinforcement learning. It uses Exponential Stability Labels and a λ-weighted aggregation to learn Lyapunov certificates, guiding policy optimization with a stability-aware advantage function, ensuring rapid Lyapunov descent and robust state convergence, outperforming baselines in stabilization and tracking tasks.
256. PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers
Core Problem: The quadratic complexity of attention in Diffusion Transformers bottlenecks efficiency for video and image generation, especially at high sparsity where context is lost, leading to degradation.
Key Innovation: PISA, a training-free Piecewise Sparse Attention, introduces an exact-or-approximate strategy that maintains exact computation for critical blocks while efficiently approximating the remainder through block-wise Taylor expansion, achieving sub-quadratic complexity and bridging the gap between speed and quality.
257. MiTA Attention: Efficient Fast-Weight Scaling via a Mixture of Top-k Activations
Core Problem: Scaling fast weights in Transformer attention becomes prohibitively expensive for extremely long sequences, limiting the expressive capacity of the N-width MLP representation.
Key Innovation: MiTA Attention proposes a compress-and-route strategy that compresses the N-width MLP into a narrower one using landmark queries and constructs deformable experts by gathering top-k activated key-value pairs for each landmark query, enabling efficient fast-weight scaling.
258. ECORE: Energy-Conscious Optimized Routing for Deep Learning Models at the Edge
Core Problem: Real-time vision-based analytics on resource-constrained edge devices require a joint optimization of energy consumption and detection accuracy.
Key Innovation: ECORE, a framework integrating multiple dynamic routing strategies (including novel estimation-based and greedy selection algorithms) to direct image processing requests to the most suitable edge device-model pair, reducing energy and latency while maintaining detection performance.
259. Trustworthy AI-based crack-tip segmentation using domain-guided explanations
Core Problem: Ensuring the trustworthiness, robustness, and generalization capabilities of deep learning models, particularly in high-stakes scientific applications like semantic crack tip segmentation, where physically meaningful explanations are crucial.
Key Innovation: An 'attention-guided training' framework that combines explainable AI with quantitative evaluation and domain-specific priors (e.g., Williams' analytical solution) to guide model attention, enhancing generalization and producing more faithful explanations for crack tip segmentation in materials.
260. AWaRe-SAC: Proactive Slice Admission Control under Weather-Induced Capacity Uncertainty
Core Problem: Millimeter-wave (mmWave) links' susceptibility to weather-related attenuation creates uncertainty about future network capacity, making it challenging to make admission control decisions for slices with QoS requirements while balancing rewards and QoS-violation risks.
Key Innovation: Develops AWaRe-SAC, a proactive slice admission control framework that integrates a predictor for short-term attenuation and uncertainty quantification with an admission control algorithm to maximize rewards and minimize QoS-violation penalties.
261. Beyond Predictive Uncertainty: Reliable Representation Learning with Structural Constraints
Core Problem: Uncertainty estimation in machine learning traditionally focuses only on the prediction stage, implicitly assuming learned representations are deterministic and reliable, neglecting representation-level uncertainty.
Key Innovation: Proposes a principled framework for reliable representation learning that explicitly models representation-level uncertainty and leverages structural constraints as inductive biases to regularize the space of feasible representations, encouraging stability and robustness.
262. Information-Theoretic Causal Bounds under Unmeasured Confounding
Core Problem: Existing approaches for identifying causal effects under unmeasured confounding often rely on restrictive assumptions, require external inputs, necessitate full structural causal model specifications, or focus only on population-level averages.
Key Innovation: Develops a data-driven information-theoretic framework that establishes novel divergence bounds, enabling sharp partial identification of conditional causal effects directly from observational data without restrictive assumptions or external parameters.
263. Bias-Reduced Estimation of Finite Mixtures: An Application to Latent Group Structures in Panel Data
Core Problem: Maximum likelihood estimation of finite mixtures of parametric densities can suffer from substantial finite-sample bias in all parameters, especially with outliers or high overlap among mixture components.
Key Innovation: Shows that maximizing the classification-mixture likelihood function, equipped with a consistent classifier, yields parameter estimates that are less biased than standard MLE and derives its asymptotic distribution, outperforming MLE in finite samples.
264. MapDream: Task-Driven Map Learning for Vision-Language Navigation
Core Problem: Most existing Vision-Language Navigation (VLN) approaches rely on hand-crafted map representations constructed independently of the navigation policy, which may not optimally aggregate spatial context for navigation objectives.
Key Innovation: Proposes MapDream, a map-in-the-loop framework that formulates map construction as autoregressive bird's-eye-view (BEV) image synthesis, jointly learning map generation and action prediction to distill navigation-critical affordances.
265. Optimal Decision-Making Based on Prediction Sets
Core Problem: While prediction sets can provide guaranteed coverage for unknown test outcomes, it remains unclear how to use them optimally for downstream decision-making to minimize expected loss.
Key Innovation: Proposes a decision-theoretic framework that minimizes the expected loss against a worst-case distribution consistent with the prediction set's coverage guarantee, and introduces Risk-Optimal Conformal Prediction (ROCP) for practical implementation.
266. Online Fine-Tuning of Pretrained Controllers for Autonomous Driving via Real-Time Recurrent RL
Core Problem: Deploying pretrained policies in real-world autonomous driving applications faces substantial challenges as fixed policies rapidly degrade in performance when encountering environmental changes in system dynamics, sensor drift, or task objectives.
Key Innovation: Shows that Real-Time Recurrent Reinforcement Learning (RTRRL), a biologically plausible algorithm for online adaptation, can effectively fine-tune a pretrained policy to improve autonomous agents' performance on driving tasks, synergizing with a recurrent network model.
267. Novel anisotropic elastoplastic modeling of K0-consolidated sands
Core Problem: Existing constitutive models for K0-consolidated sands struggle to accurately define the yield surface and capture both inherent and induced anisotropy, limiting their predictive accuracy for plastic deformation under various stress paths.
Key Innovation: Development of a novel elastoplastic constitutive model for K0-consolidated sands, incorporating new anisotropic yield functions with flexible shape control and an anisotropic state variable within a bounding surface plasticity framework, which significantly improves predictive accuracy by capturing inherent and induced anisotropy.
268. A hybrid model based on a dual-layer decomposition framework and LSTM-Informer for significant wave height prediction
Core Problem: Conventional models underperform in significant wave height (SWH) forecasting due to the non-linearities and non-smoothness of ocean waves, which is crucial for ocean engineering and maritime safety.
Key Innovation: Proposed a hybrid model (SMD-LSTM-Informer) combining STL decomposition, variational mode decomposition (VMD), LSTM, and Informer, demonstrating superior performance in both short- and long-term SWH forecasting across multiple buoy stations.
269. Motion prediction enhancement for offshore platforms under typhoon conditions using DynaSparse-ConvLSTM with dynamic sparse gating and multi-source fusion
Core Problem: Predicting offshore platform motion responses under typhoon conditions is challenging due to complex spatiotemporal coupling, underutilized motion periodicity, and limitations in real-time performance.
Key Innovation: Proposed DynaSparse-ConvLSTM architecture with multi-source feature fusion, a hierarchical feature-enhancement network, and a dynamic sparse gating mechanism, achieving significantly lower prediction errors and strong generalization capability for offshore platform motion under typhoon conditions.
270. A staged calibration strategy for SWAN whitecapping term coefficients to improve significant wave height and mean wave period
Core Problem: The SWAN model, while widely used for wind-wave simulation, frequently underestimates the mean wave period (Ta) despite accurate significant wave height (Hs), limiting its accuracy for ocean structure design and safety.
Key Innovation: Developed a two-stage calibration strategy to optimize SWAN model whitecapping term coefficients, achieving simultaneous and significant improvement in both significant wave height (Hs) and mean wave period (Ta) predictions in coastal estuaries, resolving persistent Ta underestimation.
271. Experimental investigation of hydrodynamic characteristics of deep-sea mining riser subjected to vortex-induced vibration
Core Problem: Accurately characterizing the hydrodynamic load distribution and vortex-induced vibration (VIV) of deep-sea mining risers under uniform flow, which is crucial for structural design and preventing fatigue failure.
Key Innovation: A novel inverse identification approach based on discrete strain responses and a modified Euler-Bernoulli beam equation was developed to determine hydrodynamic loads, revealing that VIV generates non-uniform, periodic loads, amplifies mean drag, and causes asymmetric in-line vortex loads in mining risers, unlike conventional risers.
272. A wind load assessment method for semi-submersible platforms based on wind profile conversion
Core Problem: Simulating NPD wind profiles for offshore platform tests is time-consuming and complex, hindering efficient wind load assessment for semi-submersible platforms.
Key Innovation: A wind load assessment method based on wind profile conversion is proposed, allowing precise prediction of wind loads for various NPD profiles from a single Power-Law test, significantly reducing simulation time and enhancing the efficiency of wind tunnel experiments for offshore platforms.
273. OPB and IPB phenomena on mooring chain observed by scaled dynamic tank experiment
Core Problem: The direct contribution of floater dynamics to Out-of-Plane Bending (OPB) and In-Plane Bending (IPB) and subsequent fatigue damage in mooring chains is unclear due to limitations of static displacement tests.
Key Innovation: A scaled dynamic tank experiment successfully measured OPB and IPB strain under dynamic catenary motion, identifying OPB/IPB mechanisms and proposing new formulas to estimate stresses directly from floater motion, providing crucial insights for mooring system design.
274. Design, fabrication and multiple failure behavior of carbon fiber composite circumferential corrugated reinforced cylindrical shells under hydrostatic pressure
Core Problem: Designing lightweight deep-sea pressure shell structures requires accurate prediction of multiple failure behaviors (overall buckling, local buckling, strength failure) under hydrostatic pressure, which existing methods may not fully address.
Key Innovation: A carbon fiber composite circumferential corrugated reinforced cylindrical shell structure is proposed, along with a multiple failure theory that effectively predicts overall buckling, local buckling, and strength failure modes with errors generally within 10% compared to simulations, offering a novel method for lightweight deep-sea pressure shell design.
275. Enhancing prediction performance of experimental pressure on square cylinders with CFD data and deep transfer learning
Core Problem: Accurate estimation of fluid-induced loading on bluff bodies from wind tunnel experiments is often limited by insufficient sensor data.
Key Innovation: A TL-POD-LSTM framework, combining CFD data with deep transfer learning, significantly enhances the prediction performance of experimental pressure on square cylinders, especially with insufficient training data, reducing RMS errors by up to 68% and accurately predicting aerodynamic coefficients and pressure time-series.
276. Study on the characteristics and prediction model of gas-liquid two-phase slug flow in subsea jumper
Core Problem: Gas-liquid two-phase slug flow in subsea jumpers intensifies pressure fluctuations and energy losses, threatening structural integrity and operational safety, and requires better understanding and predictive models.
Key Innovation: Integrating experiments and numerical simulations, this study quantified slug flow characteristics and hydrodynamic loads in an M-shaped jumper, developing predictive correlations for slug velocity, length, and frequency with errors below 10%, thereby improving understanding for design optimization and safe multiphase transport.
277. Ultimate strength assessment of I-type sandwich panel considering coupled buckling and weld fracture failure
Core Problem: Accurately predicting the ultimate bearing capacity and failure mechanisms of ship grillages under compressive loads, particularly considering coupled buckling and weld fracture, is crucial for hull girder strength and lightweight design.
Key Innovation: A coupled shell-solid model integrating an extended NH-GTN damage model for weld elements is developed, accurately reproducing the "overall buckling followed by weld cracking" failure evolution in I-type sandwich panels and providing a quantitative basis for ultimate load capacity evaluation and connection optimization.
278. Analytical model for static behavior of modular floating structures (MFSs) with arbitrary floater numbers and sizes
Core Problem: A simple and effective static design model for modular floating structures (MFSs) with arbitrary floater numbers and sizes is lacking, hindering early-stage design for applications like floating cities.
Key Innovation: An analytical model is developed to evaluate the static behavior (deflection, inclination, connector forces) of finite MFSs by superposing infinite and semi-infinite MFSs, providing an accurate and efficient tool for early design stages of floating cities and bridges.
279. Dynamic analysis of the flexible conveying pipe in a new seabed levelling system
Core Problem: The dynamic responses of flexible conveying pipes in seabed levelling systems are critical for operational efficiency and safety in deep-water coastal construction, but the effects of environmental and operational parameters are not fully quantified.
Key Innovation: This study quantified the effects of robot movement, environmental loads, and internal flow concentration on the safety of flexible pipes in a seabed levelling system, providing essential engineering guidelines, such as recommending an internal flow concentration below 15% as an operational limit.
280. Competing vortex shedding modes in vortex-induced vibration dynamics of rectangular cylinders: geometric dependence of leading/trailing edge synchronization
Core Problem: Understanding the complex vortex-induced vibration (VIV) dynamics of rectangular cylinders, particularly the competing vortex shedding modes and their geometric dependence, is crucial for optimizing marine structures against flow-induced vibrations.
Key Innovation: DNS simulations revealed two distinct VIV lock-in ranges governed by trailing-edge and leading-edge vortex shedding, demonstrating how geometric modifications (elliptical leading edge, trailing-edge modification) critically alter VIV dynamics, providing insights for structural optimization.
281. Theoretical analysis of the local mechanical characteristics of dynamic submarine cables under axisymmetric loads
Core Problem: A lack of theoretical models and comprehensive analyses hinders the understanding of local mechanical characteristics of dynamic submarine cables (DSCs) under axisymmetric loads, which is essential for guiding structural design.
Key Innovation: A theoretical model based on energy conservation is developed to analyze DSCs under axisymmetric loads, accurately predicting local mechanical characteristics and revealing the crucial role of helical armor layers, providing recommendations for optimal design parameters.
282. TED: A global temperature-driven thermoelastic displacement dataset for GNSS reference stations (2000–2023)
Core Problem: A globally consistent and reproducible data product for temperature-driven thermoelastic deformation (TED) in GNSS station height time series has long been lacking, contributing to nonlinear signals.
Key Innovation: Presented a global dataset of vertical TED for ~15,000 GNSS stations (2000–2023), generated using a full-spectrum, layered finite-element model driven by hourly ERA5 soil-temperature profiles and SoilGrids thermophysical properties, showing that TED corrections reduce residual vertical dispersion by up to ~70% at selected sites.
283. Simulating the recent drought-induced mortality of European beech (Fagus sylvatica L.) and Norway spruce (Picea abies L.) in German forests
Core Problem: Understanding and accurately simulating drought-induced tree mortality in German forests, particularly for European beech and Norway spruce, by integrating the predisposing-inciting framework with factors like soil properties and bark beetle damage, without relying on calibration.
Key Innovation: Development and application of ForClim v4.2, a process-based model incorporating a predisposing-inciting framework and a bark beetle module, to simulate drought-related tree mortality across German forests, demonstrating the critical role of soil water holding capacity and heterogeneity, and the improved performance with the bark beetle submodel.
284. Zooming in: SCREAM at 100 m using regional refinement over the San Francisco Bay Area
Core Problem: Pushing global climate models to large-eddy simulation (LES) scales over complex terrain has remained a major challenge, limiting their ability to realistically capture fine-scale atmospheric processes and coastal phenomena.
Key Innovation: First known implementation of a global model (SCREAM) at 100 m horizontal resolution using a regionally refined mesh (RRM) over the San Francisco Bay Area, demonstrating stable LES-scale operation, realistic capture of topography, surface heterogeneity, and coastal processes, and substantial improvements in near-surface wind speed, temperature, humidity, and pressure biases compared to lower resolution simulations.
285. Operational numerical weather prediction with ICON on GPUs (version 2024.10)
Core Problem: Numerical weather prediction and climate models require continuous adaptation to leverage advances in high-performance computing hardware, specifically porting to GPUs, while meeting strict requirements for time-to-solution and meteorological quality for operational applications.
Key Innovation: Successful port of the ICON model to GPUs using OpenACC compiler directives for operational numerical weather prediction, achieving a 5.5 × speed-up compared to the CPU baseline through performance tuning and mixed-precision optimization, making MeteoSwiss the first national weather service to run ICON operationally on GPUs.
286. An original approach combining biogeochemical signatures and a mixing model to discriminate spatial runoff-generating sources in a peri-urban catchment
Core Problem: Limited application of hydrograph separation to spatially decompose flow and identify runoff-generating sources, especially in peri-urban catchments.
Key Innovation: Developed an original approach combining biogeochemical signatures and a Bayesian mixing model to spatially discriminate runoff-generating sources, improving hydrological models and anticipating influences of land use/climate change on runoff.
287. The general formulation for mean annual runoff components estimation and their change attribution
Core Problem: Lack of a general framework to quantify and attribute mean annual runoff components.
Key Innovation: Proposed a general formulation (MPS model) based on the two-stage Ponce-Shetty model to characterize and attribute mean annual runoff components, demonstrating its effectiveness across large catchments and providing insights into surface flow and baseflow responses to available water and environmental factors.
288. High resolution monthly precipitation isotope estimates across Australia from machine learning
Core Problem: Water isotope observations are sparse, limiting understanding of spatial and temporal water isotope variability and their application in water cycle studies.
Key Innovation: Developed and applied a machine learning (random forest) approach to predict high-resolution, spatially continuous monthly precipitation isotope estimates across Australia, providing a robust dataset for hydrology and palaeoclimate research.
289. Resilience assessment and enhancement of urban transportation interdependent network under cascading failure
Core Problem: Insufficient attention to response mechanisms during disturbances and post-disturbance resilience enhancement in urban transportation interdependent networks (UTINs), particularly under cascading failures.
Key Innovation: A cascading failure model that considers passenger transfer impedance and a recovery priority strategy for failed nodes to maximize UTIN resilience, providing insights into optimal transfer distances, flow distribution optimization, and effective recovery strategies for disaster prevention and emergency response.
290. A Data-Driven Risk Prediction Framework based on Multi-Criteria Sorting considering Interaction
Core Problem: Improving the accuracy, reliability, and interpretability of risk prediction in high-risk domains by accounting for multi-order feature interactions, which are often overlooked by conventional models.
Key Innovation: A data-driven risk prediction framework (SI-MCDM) integrating SHapley Interaction Quantification (SHAP-IQ) with Multi-Criteria Decision Making (MCDM) and Choquet Integral, demonstrating improved prediction performance by considering second-order failure mode interactions.
291. Cognition-inspired multimodal attention fusion of close-range laser scanning data for globally representative tree species classification
Core Problem: Species-level tree identification remains a major challenge in global environmental monitoring and AI-driven ecological assessment due to high species diversity, structural plasticity, and variability across sensing platforms.
Key Innovation: A cognition-inspired multimodal attention fusion network (CI-MAFusion), a dual-branch deep learning framework integrating point cloud data with multi-view imagery, achieving high accuracy in globally representative tree species classification by adaptively fusing structural and visual features.
292. Diversity, composition, networks, and assembly processes of soil microbial communities across slope positions in a karst peak cluster depression
Core Problem: Existing studies have focused on how slope position influences microbial diversity or composition, but ignored its impact on microbial networks and community assembly processes in karst peak cluster depressions.
Key Innovation: Investigated how slope position affects soil microbial diversity, composition, network structures, and assembly mechanisms in a karst peak cluster depression, revealing significant decreases in bacterial alpha diversity from bottom to top, distinct community structures, and varying network complexities and assembly processes, demonstrating spatial variation and potential for ecological restoration.
293. Fine characterization of micro-nano fractures and analysis of network connectivity: Mechanistic controls on hydrocarbon enrichment in shale reservoirs
Core Problem: The developmental characteristics and controlling factors of micro-nano fractures, crucial for connecting nanopores and macro-fractures, are unclear, hindering efficient development of continental shale oil.
Key Innovation: An entropy weight method established an evaluation model for fracture development strength, and topology was introduced to evaluate connectivity. The study clarified differences in micro-nano fracture development among lithofacies, showing that high organic matter content correlates with greater fracture strength, clay minerals control nano-fracture development, and felsic minerals influence connectivity. Micro-nano fractures enhance pore structure and connectivity, impacting hydrocarbon accumulation and hydraulic fracturing effectiveness.
294. Impact of vegetation greening on runoff under climate change in the Yarlung Tsangpo-Brahmaputra River basin
Core Problem: Evaluating the possible impact of vegetation greening under the CMIP6 SSP585 scenario on the runoff volume till 2100 in the Yarlung Tsangpo-Brahmaputra River basin.
Key Innovation: Prediction that vegetation greening will decrease annual runoff (3 to 31 billion m3) at Bahadurabad, with water yield decline in upper and middle reaches, highlighting the importance of accurately predicting vegetation spatial distribution and temporal dynamics in hydrological modeling for water resource management.
295. Non-uniform suspended sediment transport in turbulent open channel flows under non-equilibrium conditions
Core Problem: Investigating steady-state non-equilibrium transport for non-uniform sediment particles in turbulent open-channel flows, specifically how boundary conditions govern vertical concentration distribution and longitudinal development.
Key Innovation: An analytical framework that systematically evaluates generalized and specialized bottom boundary conditions, showing how discrepancies between deposition and entrainment rates induce flux-equilibrium states, and how particle settling velocity and non-uniformity influence vertical sediment distribution, advancing predictive modeling for riverine management.
296. Quantifying the uncertainty contribution in runoff projection and the time scale effects
Core Problem: Quantifying the uncertainty contributions of different factors (modeling chain, internal variability) to hydrological projections and exploring how these contributions respond to changes in the hydrological forecasting period.
Key Innovation: Using time-series analysis of variance, the study shows that for short-term forecasts (<20 years), internal variability dominates uncertainty (>50%), while for long-term forecasts (75 years), the modeling chain becomes the main source (80-87%), with specific contributions from GCMs and forecast models varying by basin.
297. Parameterization of complex geological models with PCA‑guided adversarial diffusion for ensemble data assimilation
Core Problem: Data assimilation of complex geological models (e.g., channelized models) in subsurface flow is challenging due to non-Gaussian characteristics and high dimensionality, requiring effective parameterization to preserve plausible geological features while reducing complexity.
Key Innovation: Proposal of principal component analysis-guided adversarial diffusion (PCA-GAD), a single-step diffusion-based generative framework combining PCA with adversarial training, to parameterize complex geological models, which preserves key geological structures, reduces dimensionality, and dramatically shortens sampling time for faster ensemble-based data assimilation.
298. Climate warming-induced glacier mass loss driving peak runoff variability and cryosphere service value decline
Core Problem: Climate warming is accelerating glacier melting in High Mountain Asia, threatening water resource sustainability and leading to a decline in cryosphere service value.
Key Innovation: Employed an integrated ice-dynamic model to reconstruct glacier mass balance, assess long-term dynamics under climate scenarios, and quantify ecological value loss from glacier retreat, providing insights into glacier hydrological responses and informing water resource management.