TerraMosaic Daily Digest: Feb 22, 2026
Daily Summary
The February 22 corpus (126 selected studies from 938 deduplicated candidates) shows a clear shift from static susceptibility mapping to process-resolving, decision-oriented hazard intelligence. High-impact landslide work combines mechanism-guided displacement forecasting, seismic damage evolution analysis, and high-resolution terrain observation (InSAR and UAV DEMs) to explain when and why slopes accelerate, not only where failures may occur. In parallel, flood and seismic studies increasingly connect hazard, exposure, and infrastructure response: building-scale flood loss quantification, road-disruption chronologies, and directivity-aware seismic hazard models translate geoscience outputs into operational risk metrics. A third line of progress is infrastructure-focused geomechanics, where reliability analysis and multi-physics simulation are used to constrain uncertainty for tunnels, offshore foundations, and utility corridors. Lower-relevance AI papers are mostly methodological, but they still point to transferable advances in uncertainty quantification, causal inference, and data-efficient modeling that can strengthen geohazard systems when grounded in physics and field evidence.
Key Trends
The strongest signals in this cycle are methodological convergence and stronger links between physical process understanding and deployable risk management.
- Mechanism-informed prediction is replacing purely static hazard mapping: Several top-ranked studies move beyond susceptibility zoning by integrating deformation mechanisms, seismic loading history, and temporal kinematics, improving predictive credibility for accelerating slopes and reactivation-prone terrain.
- Observation pipelines are becoming multi-sensor and operational: InSAR, UAV DEMs, LiDAR point clouds, and satellite optical products are increasingly fused in automated workflows, yielding higher spatiotemporal resolution and more reproducible post-event and long-term monitoring products.
- Hazard outputs are being translated into actionable impact metrics: Road interruption chronologies, building-scale flood damage quantification, and public risk-perception studies indicate a maturing interface between hazard science and infrastructure/public decision-making.
- Reliability-aware engineering is expanding across geotechnical systems: From scour protection and offshore cables to tunnel deformation and soft-rock rheology, studies increasingly combine probabilistic inference with coupled numerical models to capture failure uncertainty rather than single deterministic scenarios.
- Climate controls are treated as dynamic hazard modulators, not background context: Work on glacier change, intense wind gusts, regional drought dynamics, and Tibetan Plateau soil-rainfall teleconnections shows stronger emphasis on climate-conditioned triggering pathways and seasonal predictability.
Selected Papers
This digest features 126 selected papers from 938 papers analyzed (out of 2637 raw papers scanned; 938 new papers after deduplication) 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. Mechanism guided forecasting model for landslide displacement prediction: case study for Baishuihe Landslide
Core Problem: Accurate early warning for landslides is challenging due to step-like deformation and abrupt displacement increases, requiring better understanding of mechanisms and effective feature extraction.
Key Innovation: Proposes a hybrid forecasting model (VMD, DES, Informer with multi-head attention) guided by landslide evolution mechanisms, achieving high prediction accuracy for Baishuihe Landslide displacement.
2. Seismic damage evolution at Yigong mountain driven by the repeated earthquake swarms along Jiali Fault in Eastern Himalayan Syntaxis
Core Problem: Understanding the mechanical mechanisms of progressive failure in rock slopes under repeated seismic loading, specifically how it led to the massive 2000 Yigong landslide in a high-intensity seismic zone.
Key Innovation: Revealed a three-stage progressive failure mode (initial damage, stable deformation, final instability) under seismic conditions using field surveys and discrete element numerical simulations, identifying rock mass fragmentation and strength degradation from frequent earthquakes as key triggering factors.
3. Rock slope failures in Baspa Valley, Himachal Pradesh, India: a multi method numerical assessment and mitigation strategies
Core Problem: Rock slope instability along the Sangla–Chitkul corridor in the Higher Himalayas constitutes a significant hazard, requiring robust numerical assessment and effective mitigation strategies.
Key Innovation: Developed a comparative numerical modeling framework (DEM and DFN–FEM) to evaluate rock slope stability, identifying critical geometric and geomechanical drivers of instability, and demonstrating the effectiveness of geometric re-profiling and rock bolting for mitigation.
4. Road interruptions due to landslides and floods in Southern Italy
Core Problem: Road networks in regions exposed to rainfall-induced geo-hydrological hazards are highly vulnerable to service disruptions, yet systematic analyses of road interruption data remain limited.
Key Innovation: Conducted a systematic analysis of road closures caused by landslides and floods in Calabria, revealing distinct hazard signatures (frequency, duration, spatial patterns) and providing insights for road managers and a reproducible framework for analyzing road interruption chronologies.
5. Consideration of near-field directivity effects in probabilistic seismic hazard assessment: a case study over North–East India
Core Problem: Traditional probabilistic seismic hazard assessments often overlook near-field directivity effects, leading to less accurate hazard estimates for critical infrastructure in close proximity to fault ruptures.
Key Innovation: Developed a PSHA for North–East India that incorporates near-field directivity effects using a logic tree approach, demonstrating significant amplification and spatial variation of these effects, thereby improving the accuracy of seismic hazard estimates.
6. A Novel Data‐Driven Probabilistic Approach for Consequence Specific Urban Flood Risk Mapping
Core Problem: Conventional hydrologic and hydraulic models for urban flood-prone areas demand extensive data and computation, limiting their use, while existing risk mapping approaches often rely on subjective judgments for vulnerability indicator weighting.
Key Innovation: Develops a data-driven probabilistic framework for consequence-specific urban flood risk mapping, integrating hazard and vulnerability probabilities, and introduces an Empirical-Improved Analytical Hierarchy Process (E-IAHP) to objectively derive indicator weights using statistical measures, enabling tailored risk maps for specific consequences.
7. Decadal glacier displacement and emerging hazards in Sikkim, Eastern Himalaya from InSAR observations (2014–2024)
Core Problem: Lack of long-term monitoring of glacier thinning and dynamics in the Sikkim Himalayas, which increases the risk of glacial lake outburst floods (GLOFs) and other related hazards in downstream areas.
Key Innovation: Presented a ten-year evaluation of glacier dynamics in Sikkim using InSAR, revealing widespread thinning, specific high-loss glaciers (e.g., South Lhonak contributing to 2023 GLOF), surge-like signals, and varied responses by morphology, integrating climate data for improved hazard prediction.
8. Developing predictive models to assess coal burst phenomena using machine learning algorithms
Core Problem: Coal burst is a significant geohazard in underground coal mines, and accurate prediction of its liability is crucial for preventing and mitigating its destructive impacts.
Key Innovation: Developed and validated five white-box machine learning models, particularly a tree-augmented naive Bayes model, for accurate prediction of coal burst liability using key variables, and utilized SHAP analysis for model interpretability and variable contribution assessment.
9. Enhancement of the crack and erosion resistance of silty clay under freeze–thaw cycles: Synergistic effect of sisal fiber–fly ash
Core Problem: Freeze–thaw (F–T) cycle-induced cracking in silty clays poses a significant risk to engineering stability, and individual soil improvement methods offer only partial solutions.
Key Innovation: Demonstrated the synergistic effect of sisal fiber (SF) and fly ash (FA) in significantly enhancing the crack resistance, erosion resistance, and strength of silty clay under F-T cycles. Provided a scientific basis for designing soil improvement in disaster mitigation engineering in seasonally frozen soil regions.
10. Performance of SRTM, PALSAR, and UAV DEMs in identifying landslides: Akchour, Morocco
Core Problem: Precise examination of the Akchour landslide is hindered by its steep and extensive nature, and traditional broad-scale aerial photos lack the detail needed for local-scale geomorphological and geomorphometric analysis.
Key Innovation: A comparative study demonstrating the superior potential of UAV-generated high-resolution DEMs (17 cm ground resolution) over freely available SRTM and PALSAR DEMs for detailed geomorphological study, orthophoto generation, and understanding dynamic behaviors and potential risks within the Akchour landslide, especially in inaccessible areas.
11. Improving Landslide Susceptibility Prediction Through Hybrid Framework: A Case from the Deccan Province
Core Problem: Landslide susceptibility assessment in basaltic terrains is challenging due to complex geomorphic controls and limited evaluation of integrated modeling approaches.
Key Innovation: Developed a hybrid modeling framework integrating Modified Frequency Ratio (MFR) and Analytic Hierarchy Process (AHP) models, which significantly improved landslide susceptibility prediction (AUC of 0.827) compared to individual models, providing a robust tool for regional hazard assessment.
12. Development of a building-scale integrated flood damage quantifying framework using a hydrodynamic model and multisource geospatial data
Core Problem: The need for a precise and comprehensive framework to quantify flood-induced damage to individual building properties (structural and content), overcoming limitations of traditional survey methods for effective disaster risk reduction and community resilience planning.
Key Innovation: A comprehensive framework integrating geospatial data, machine learning (random forest for building classification achieving 98.4% accuracy), and 2D hydrodynamic modeling to quantify building-scale flood damage. Applied to a real flood event, it provided detailed damage ratios and economic loss estimates, offering a transferable approach for disaster resilience.
13. Augmenting seismic sustainability of unreinforced brick masonry walls through cost-efficient nominal RC band interventions
Core Problem: Non-engineered unreinforced masonry walls, common in many regions, are highly susceptible to seismic damage, requiring cost-effective retrofitting solutions.
Key Innovation: Investigated a novel, cost-efficient retrofitting technique using nominal reinforced concrete bands within wall thickness, demonstrating significant improvement in seismic performance and structural integrity through shake table tests and fragility analysis, with a low estimated cost.
14. Effects of liquefiable soil spatial distribution on seismic dynamic response of two-story subway stations
Core Problem: The spatial distribution of liquefiable soils significantly affects the seismic performance of subway stations, but its coupled effects with ground-motion characteristics on structural response are not fully understood.
Key Innovation: Employed finite element method and nonlinear dynamic time-history analyses to systematically elucidate the coupled effects of liquefiable-layer location and ground-motion spectral characteristics on subway station seismic response, finding that liquefiable sites markedly increase inter-story displacement and that the case with liquefiable layers on both sides is most critical.
15. Satellite‐Based Estimation of Near‐Surface Specific Humidity During Tropical Cyclones in the Western North Pacific and North Atlantic
Core Problem: Accurately estimating near-surface specific humidity during tropical cyclones is challenging for satellite microwave radiometers due to heavy precipitation masking the water-vapor signal, which hinders accurate enthalpy flux estimation and understanding of storm intensification.
Key Innovation: Introduced a two-pronged approach using a neural network for precipitation-contaminated measurement screening and the first use of low-frequency microwave channels for humidity estimation in tropical cyclones, leading to more accurate near-surface humidity estimates comparable to non-storm conditions.
16. Comparative Assessment of Multimodal Earth Observation Data for Soil Moisture Estimation
Core Problem: Existing satellite soil moisture (SM) products are too coarse (>1km) for farm-level applications, and there's a need to optimize the combination of multimodal Earth Observation data for high-resolution SM estimation.
Key Innovation: Presents a high-resolution (10m) SM estimation framework for vegetated areas across Europe, combining Sentinel-1 SAR, Sentinel-2 optical, and ERA-5 data via machine learning, demonstrating that hybrid temporal matching and traditional feature engineering offer a practical solution for pan-European field-scale soil moisture monitoring.
17. A Deep Surrogate Model for Robust and Generalizable Long-Term Blast Wave Prediction
Core Problem: Accurately modeling long-term spatio-temporal blast wave propagation is challenging due to its nonlinearity, sharp gradients, computational cost, and the degraded accuracy and error accumulation of existing ML surrogate models on complex or out-of-distribution scenarios.
Key Innovation: Proposing RGD-Blast, a deep surrogate model that incorporates a multi-scale module to capture global flow patterns and local boundary interactions, and a dynamic-static feature coupling mechanism for enhanced out-of-distribution generalization, achieving high accuracy and a two-order-of-magnitude speedup over traditional methods.
18. Reliability-based dynamic damage modelling for monopile scour protection under combined waves and currents
Core Problem: Traditional deterministic models for scour around monopile foundations fail to capture inherent uncertainties and non-linear dynamics, hindering accurate risk assessment and design of protection systems.
Key Innovation: A probabilistic framework using Model Tree, Bayesian Regression, and Explanatory Regression (ER) models to evaluate dynamic scour-induced damage, quantifying damage evolution probabilistically and offering enhanced insight for risk-informed design and maintenance of offshore infrastructure.
19. Investigations on alleviation effect of Bragg breakwater on harbor resonance induced by irregular waves
Core Problem: Harbor resonance induced by irregular waves poses a risk to coastal infrastructure, and understanding the optimal mitigation effects of Bragg breakwaters, especially with varying spectral peak frequencies, is crucial.
Key Innovation: A study using a fully nonlinear Boussinesq model (FUNWAVE 2.0) demonstrating that Bragg breakwaters effectively mitigate harbor resonance excited by irregular waves with varying spectral peak frequencies, identifying optimal wavelength ratios for individual modes and overall harbor resonance.
20. Using scenarios of geologic hazard events for pre-disaster mitigation
Core Problem: Understanding how scenarios of geologic hazard events can be effectively used for pre-disaster risk mitigation, given their common use but varying effectiveness in different contexts.
Key Innovation: Investigated the effectiveness of geologic hazard scenarios through cross-national interviews and literature review, identifying critical elements for effective mitigation, such as impact descriptions, action identification, collaborative development, and clear communication.
21. Flooding algorithm combining hydrology and dynamic seed growth
Core Problem: Simulating and predicting flood inundation areas accurately, particularly in data-scarce environments where traditional hydrological data limits model accuracy.
Key Innovation: Developed a flooding algorithm combining GIS, hydrology, and hydraulics with a dynamic seed propagation algorithm, demonstrating superior spatial accuracy and computational efficiency compared to mainstream models, especially in data-scarce small and medium-sized river basins.
22. Unveiling the performance of pre-processing approaches in machine learning based flood susceptibility mapping
Core Problem: Optimizing the performance of machine learning models for flood susceptibility mapping by systematically investigating the impact of various pre-processing approaches and identifying key conditioning factors.
Key Innovation: Investigated 18 pre-processing scenarios for ML-based flood susceptibility mapping using XGBoost, finding robust scaling with 70/30 train-test split and a 10x class imbalance ratio with random under sampling yielded the highest performance, and identified key conditioning factors like alluvium presence and distance to faults.
23. A study on machine learning-based prediction model for surface litter moisture content in coniferous forests
Core Problem: Accurate daily prediction of surface litter moisture content (LMC) is crucial for forest fire risk forecasting, but existing methods have limitations, especially across different LMC ranges and in systematic comparisons with traditional approaches.
Key Innovation: Developed and systematically compared machine learning models (XGBoost, Random Forest) with traditional semi-physical models for daily LMC prediction, demonstrating superior accuracy and interpretability of ML methods for forest fire prevention and fuel management.
24. Tectonic dynamics shaping the lake Hazar Basin along the East Anatolian fault system: Insights from fault kinematics and structural evolution
Core Problem: Examining transtensional deformation and the polyphase tectonic history in complex intracontinental strike-slip zones, specifically the Lake Hazar basin along the East Anatolian Fault System (EAFS).
Key Innovation: Integrated field mapping, fault-slip analysis, and focal mechanism inversion to reconstruct geologically consistent stress tensors and clarify the temporal transition from strike-slip to transtensional deformation. Highlighted the importance of discriminating polyphase fault-slip data for resolving deformation dynamics and refined understanding of strain localization and fault reactivation.
25. Multi-source validation of ecological sensitivity in the Toraja Highlands, South Sulawesi, Indonesia
Core Problem: The Toraja Highlands, a strategic watershed with steep terrain and active land-cover change, lacks an objective and validated ecological sensitivity map to support sustainable mountain watershed management and risk reduction.
Key Innovation: Construction of an ecological sensitivity index using principal component analysis (based on land cover, NDVI, slope, and rainfall) and its multi-source validation against independent landslide records (achieving 67-68% accuracy), field verification (Kappa 0.847-0.871), and stakeholder appraisal, providing a transparent workflow for identifying priority areas for reforestation, soil conservation, slope stabilization, and evidence-based risk reduction.
26. A Unified Framework for Automated Damage Assessment in Post-Disaster Built Environments Using LiDAR Point Clouds
Core Problem: The urgent need for accurate, timely, and spatially resolved post-disaster damage assessments in built environments, as traditional manual or limited data approaches fail to capture complex, localized impacts, hindering effective response and recovery.
Key Innovation: A unified, automated framework leveraging LiDAR point cloud data for detailed and quantitative damage assessment in post-disaster built environments. It involves pre-processing, segmentation refinement, and damage assessment, achieving over 95% F1-score for building identification and successfully detecting multiple structural damage types in tornado-affected areas.
27. Parametric optimization of sinusoidal curve negative Poisson’s ratio energy-absorbing structures
Core Problem: Deep tunnel support systems suffer from insufficient support resistance and low energy absorption efficiency, leading to prominent safety problems like rock burst in deep mining.
Key Innovation: A novel sinusoidal curved negative Poisson’s ratio (NPR) energy-absorbing component is proposed and parametrically optimized, achieving significantly higher support force and NPR effect, providing a new solution for rock burst prevention and control in deep mines.
28. Mapping global post- earthquake ecosystem damage boundaries
Core Problem: The delineation of ecosystem damage boundaries caused by earthquakes, beyond physical landslides, remains lacking, requiring a fully automated global mapping framework.
Key Innovation: Developed a fully automated framework using Landsat data and the continuous change detection and classification algorithm on Google Earth Engine, combined with a novel method (kernel density estimation, Otsu adaptive threshold, and morphological operations), to extract and produce the first global Post-earthquake Ecosystem Damage Boundaries (PEDBs) dataset.
29. Climatology and Environmental Controls of Intense Wind Gusts in the Brazilian Amazon
Core Problem: Lack of comprehensive, multi-decadal documentation and understanding of the climatology and environmental controls of intense convective wind gusts in the Brazilian Amazon, despite their significant impacts on forests and communities.
Key Innovation: Presented the first multi-decadal (2000–2024) assessment of intense convective wind gusts across the Brazilian Amazon, identifying their frequent occurrence, seasonal and diurnal patterns, and the thermodynamic factors (e.g., downdraft CAPE, steep low-level lapse rates) that favor their generation.
30. Solving and learning advective multiscale Darcian dynamics with the Neural Basis Method
Core Problem: Most 'physics-informed' machine learning formulations treat governing equations as a penalty loss, blurring operator structure and confounding solution approximation error with governing-equation enforcement error, making the solving and learning progress hard to interpret and control for complex multiscale dynamics.
Key Innovation: The Neural Basis Method, a projection-based formulation, couples a physics-conforming neural basis space with an operator-induced residual metric to obtain a well-conditioned deterministic minimization for advective multiscale Darcian dynamics. This method provides accurate and robust solutions in single solves and enables fast, effective parametric inference with operator learning, with a computable certificate tied to approximation and enforcement.
31. Automated machine learning-based predictive models for multi-hazard catastrophic risk assessment in offshore wind turbines
Core Problem: Finite element simulations for offshore wind turbine structural responses under multi-hazard conditions (including earthquakes) are computationally intensive and impractical for real-time risk assessment.
Key Innovation: A data-driven automated machine learning (AutoML) framework that develops rapid and reliable surrogate predictive models for offshore wind turbine structural responses under multi-hazard conditions (wind, waves, earthquakes), enabling efficient real-time risk assessment and decision support.
32. On the non-breaking solitary wave runup on a smooth slope using Synolakis solution and ballistic model
Core Problem: Existing ballistic models for non-breaking solitary wave runup on slopes do not accurately reproduce the initial acceleration of shoreline movement.
Key Innovation: A comprehensive investigation using the Synolakis (S87) nonlinear solution and a modified ballistic model (BM) that can reproduce the initial acceleration of shoreline movement, providing a more accurate description of non-breaking solitary wave uprush and its correlation with hydrostatic pressure gradient.
33. Explainable dual-point near-bit diagnostics for deepwater drilling anomalies: A deployable two-stage MI-AHP framework
Core Problem: Diminished response time for well-control anomalies in deepwater drilling due to narrow pressure margins, and signal delays/attenuation with conventional surface-only monitoring, hindering early detection of near-bit disturbances.
Key Innovation: A dual-point near-bit diagnostic framework using downhole measurements and a two-stage Multi-Indicator Analytic Hierarchy Process (MI-AHP) model to provide early anomaly warnings (kick, lost circulation, washout, and bit sticking) with high accuracy and real-time deployability.
34. Integrating multi-dimensional features for remote sensing–based drought monitoring and driver analysis during the vegetation growing season: a case study in northern Xinjiang
Core Problem: Reliable drought monitoring across heterogeneous vegetation zones in northern Xinjiang remains difficult with single-index remote-sensing indicators.
Key Innovation: Builds a multi-dimensional remote-sensing feature system and couples it with driver attribution analysis to separate climatic and surface-condition controls on growing-season drought.
35. Flood risk perception and water infrastructure: understanding public response to climate change
Core Problem: Understanding how public perception of flood risk influences trust in governmental water infrastructure management and preparedness, especially as climate change intensifies floods.
Key Innovation: Used survey data and Structural Equation Modeling (SEM) to show that higher awareness of flood risks is associated with lower trust in governmental preparedness, highlighting the impact of socio-economic conditions and underscoring the need for enhanced institutional communication and disaster preparedness.
36. Navigating climate change in carbon negative Bhutan: Insights from policy influencers and comparison to the wider Himalayan region
Core Problem: Bhutan and the wider Himalayan region face intensified climate risks, including glacial lake outburst floods, water scarcity, and agricultural disruptions, alongside governance challenges despite Bhutan's carbon-negative status.
Key Innovation: Explored how Bhutan navigates climate change through policy influencer perspectives, comparing findings with the broader Himalayan region. Identified shared vulnerabilities and governance challenges, advocating for integrated policies, stronger regional cooperation, and equitable climate financing.
37. Tracing sediment sources during rainfall events in a northern Loess Plateau catchment using geochemical and mid-infrared spectral methods
Core Problem: Accurate identification of sediment sources and erosion hotspots is crucial but challenging for effective soil and water conservation in highly erodible catchments.
Key Innovation: Developed and validated mid-infrared (MIR) spectroscopy as a reliable, non-destructive, and cost-effective alternative to geochemical fingerprinting for rapid sediment source apportionment, demonstrating its effectiveness in a highly erodible Loess Plateau catchment.
38. Effects of Soil Creep on the Excavation Induced Responses of Existing Tunnels in Soft Ground
Core Problem: The long-term safety of metro tunnels adjacent to deep excavations in soft soil regions is significantly limited by soil rheological (creep) behavior, leading to challenges in accurate deformation prediction.
Key Innovation: Investigated the multiscale creep mechanism of Fuzhou mucky soft soil and its effect on the soil–pit–tunnel system, demonstrating that incorporating creep effects into finite element models significantly improved tunnel deformation prediction accuracy (reducing errors from 12.5–15.8% to 2.5–6.7%) and proposed an optimized support system.
39. Peridynamics model and application for shield tunnel segments coupling beam-spring and Winkler foundation models
Core Problem: Unreasonable backfill grouting parameters in shield tunnels can cause segment cracking and reduced bearing capacity, compromising tunnel operation safety.
Key Innovation: A coupled mechanical model integrating beam-spring and Winkler foundation models into a state-based peridynamics system is constructed to simulate segment-foundation interaction and analyze the effect of grouting pressure on segment mechanical response, recommending stepwise and zonal grouting to enhance safety.
40. Pressure transfer patterns under different conditions of filter cake formation for slurry shield tunnelling
Core Problem: Understanding the mechanisms of pressure transfer by filter cake is crucial for maintaining tunnel face stability during slurry shield tunnelling in saturated sand, but individual differences among transfer patterns are not well understood.
Key Innovation: The mechanisms of pressure transfer under different filter cake formation conditions (external, internal infiltrated zone, or both) are investigated, showing how pore pressure decreases differently based on bentonite concentration and filter cake type, providing guidance for adjusting slurry pressure to enhance tunnel face stability.
41. Long-term mechanical behavior of load-shedding culverts on soft soil foundation
Core Problem: The long-term stress characteristics and load reduction efficacy of load-shedding culverts on soft soil foundations remain uncertain, impacting their design and performance prediction.
Key Innovation: Theoretical methods and numerical simulations are used to analyze the short-term and long-term earth pressures around load-shedding culverts, revealing how foundation soil consolidation and EPS creep affect vertical and lateral earth pressures and foundation contact pressure over time, providing a valuable reference for design and long-term performance prediction.
42. Effects of particle surface roughness on the mechanical behaviour of unsaturated granular materials: DEM simulations
Core Problem: Most existing DEM simulations for unsaturated granular materials assume smooth particles, leading to overestimation of capillary stresses and less realistic predictions of mechanical behavior relevant to slope stability.
Key Innovation: Implemented an enhanced capillary bridge model accounting for particle surface roughness in DEM simulations of triaxial and biaxial tests, demonstrating that roughness reduces capillary cohesion effects, making mechanical responses more closely resemble dry granular materials, and providing a more realistic approach for hydro-mechanical response investigations.
43. Enabling Training-Free Text-Based Remote Sensing Segmentation
Core Problem: Existing text-based remote sensing segmentation approaches rely on additional trainable components, limiting their generalization and practical applicability.
Key Innovation: Introduces a largely training-free pipeline that combines CLIP/SAM mask selection with VLM-generated prompts (plus lightweight LoRA tuning), delivering strong zero-shot segmentation performance across 19 remote-sensing benchmarks.
44. Causality by Abstraction: Symbolic Rule Learning in Multivariate Timeseries with Large Language Models
Core Problem: Inferring generalized and interpretable causal relations in timeseries data with delayed effects is challenging, especially for complex dynamical systems where traditional approaches fail to provide clear explanations.
Key Innovation: Presents ruleXplain, a framework that leverages Large Language Models to extract formal, verifiable causal rules with temporal operators and delay semantics from simulation-driven dynamical systems, by generating and clustering counterfactual inputs and employing a closed-loop refinement process.
45. A Single Image and Multimodality Is All You Need for Novel View Synthesis
Core Problem: Single-image novel view synthesis using diffusion models is limited by the fragility and unreliability of monocular depth estimates, especially in challenging real-world conditions (low texture, adverse weather, occlusion).
Key Innovation: Introduced a multimodal depth reconstruction framework that leverages sparse range sensing data (radar/LiDAR) to produce robust dense depth maps, significantly improving geometric consistency and visual quality in single-image novel-view video generation.
46. Causal Neighbourhood Learning for Invariant Graph Representations
Core Problem: Graph data often contains spurious correlations that hinder Graph Neural Networks (GNNs) from learning true causal relationships and generalizing effectively across different graphs, especially under distribution shifts.
Key Innovation: Proposed Causal Neighbourhood Learning with Graph Neural Networks (CNL-GNN), a framework that performs causal interventions on graph structure to identify and preserve causally relevant connections, learning invariant node representations that are robust and generalize well across different graph structures.
47. Neural-HSS: Hierarchical Semi-Separable Neural PDE Solver
Core Problem: Deep learning methods for solving Partial Differential Equations (PDEs) are constrained by the substantial computational costs associated with generating large-scale, high-quality datasets and training models, despite their ability to enable fast simulations once trained.
Key Innovation: Neural-HSS, a parameter-efficient architecture built upon the Hierarchical Semi-Separable (HSS) matrix structure, which is provably data-efficient for a broad class of PDEs, demonstrating superior ability to learn from low-data regimes and applicable to diverse domains like fluid dynamics, electromagnetism, and biology.
48. Learning Flow Distributions via Projection-Constrained Diffusion on Manifolds
Core Problem: Existing diffusion-based generative models for fluid flows either ignore physical constraints, impose soft penalties that do not guarantee feasibility, or specialize to fixed geometries, hindering the synthesis of physically accurate incompressible flows under arbitrary obstacle geometries and boundary conditions.
Key Innovation: A generative modeling framework integrates a boundary-conditioned diffusion model, a physics-informed training objective with a divergence penalty, and a projection-constrained reverse diffusion process that enforces exact incompressibility through a geometry-aware Helmholtz-Hodge operator, enabling the synthesis of physically feasible 2D incompressible flows with significantly improved accuracy and boundary consistency.
49. On the Generalization and Robustness in Conditional Value-at-Risk
Core Problem: The statistical behavior, generalization, and robustness of Conditional Value-at-Risk (CVaR) under heavy-tailed and contaminated data, particularly its dependence on an endogenous quantile, are poorly understood.
Key Innovation: Develops a learning-theoretic analysis of CVaR-based empirical risk minimization, establishing sharp generalization and excess risk bounds under minimal moment assumptions, deriving a uniform Bahadur-Kiefer type expansion for quantile sensitivity, and proposing a truncated median-of-means CVaR estimator for robustness under contamination.
50. Data-Efficient Inference of Neural Fluid Fields via SciML Foundation Model
Core Problem: Inferring neural fluid fields and realistic rendering of fluid dynamics requires dense, costly real-world captures, and the transferability of SciML foundation models to real-world vision problems is underexplored.
Key Innovation: A method leveraging SciML foundation models to significantly reduce data requirements and improve generalization for inferring real-world 3D fluid dynamics, using a collaborative training strategy that equips neural fluid fields with augmented frames and fluid features.
51. An analytical framework for torsional dynamic response of floating piles in viscoelastic unsaturated transversely isotropic soils
Core Problem: The torsional dynamic response of floating piles in complex soil conditions (viscoelastic, unsaturated, transversely isotropic) is not adequately addressed, especially considering the finite-thickness soil layer to bedrock.
Key Innovation: An analytical framework using a fictitious unsaturated soil pile (FUSP) methodology to assess the torsional dynamic response of floating piles, integrating three-phase interactions, transverse isotropy, and fractional calculus-based viscoelasticity to provide closed-form and semi-analytical solutions.
52. Dynamic coupling mechanism and parametric study of mooring chain–seabed interaction based on a global-to-local nested modeling approach
Core Problem: The complexity of dynamic interaction between anchor chains and the seabed in deepwater mooring systems makes accurate simulation challenging, despite its criticality for chain displacement and mooring anchor capacity.
Key Innovation: A 'global-to-local' nested modeling framework that establishes a coupled floater-mooring-seabed model and a refined chain-soil model (using CEL method) to analyze chain-soil interaction under realistic ocean loads, providing insights into soil resistance, deformation patterns, and trenching development.
53. Dynamic behavior of long-span cable-stayed bridge under ship-pylon collision
Core Problem: Assessing the dynamic behavior, vulnerability, and potential damage of long-span cable-stayed bridges when subjected to ship-pylon collisions.
Key Innovation: A refined finite element model and a modified mass weighted damping method to accurately simulate and analyze the global and local dynamic responses of a cable-stayed bridge under ship-pylon collision, revealing specific damage mechanisms, foundation displacements, and energy absorption characteristics.
54. Munition Piles in the German Baltic Sea: Inventory and Maritime Hazard Perspectives
Core Problem: Lack of detailed, accurate datasets on the distribution and characteristics of dumped munition piles in the German Baltic Sea, hindering effective management, risk assessment, and remediation planning.
Key Innovation: Presentation of two novel datasets detailing the distribution, properties, and geographic context of 484 known and potential munition piles in the German Baltic Sea, derived from multibeam echosounder data and photomosaics, enabling extrapolation of object numbers and identification of high-risk clusters.
55. Warning response behavior model: conceptual clarification and operationalization
Core Problem: Inadequate public response to Early Warning Systems (EWS) despite technological advancements, due to conceptual confusion and inconsistent measurement in warning dissemination and communication (WDC) research.
Key Innovation: Systematically reviewed behavioral concepts in natural hazards to propose a precise, operational framework for evaluating WDC effectiveness, covering information processing, cognitive evaluation, and intention-action conversion, supported by a longitudinal experiment.
56. Effect of doppler radar reflectivity and radial velocity assimilation on lightning and rainfall prediction of a severe thunderstorm over Odisha, India
Core Problem: Accurate prediction of severe thunderstorms, lightning occurrence, and heavy rainfall during short-duration weather events is challenging, and current models can be improved by better data assimilation.
Key Innovation: Demonstrated that assimilating Doppler Weather Radar reflectivity and radial velocity data significantly enhances the NCUM-R model's capability in simulating convective systems, leading to improved forecasts of thunderstorm activity, lightning occurrence, and heavy rainfall patterns.
57. Uncertainty Estimation of Rheological Parameters of Soft Rock Using Bayesian Inference and its Application
Core Problem: Uncertainty in soft rock rheological parameters poses significant challenges for both laboratory calibration and engineering-scale application, particularly for predicting long-term deformation in soft rock tunnels.
Key Innovation: Proposed a Bayesian uncertainty estimation framework using Markov Chain Monte Carlo (MCMC) for soft rock rheological parameters, enabling probabilistic prediction of long-term tunnel deformation that aligns with field monitoring data within a 95% confidence interval.
58. Automated demarcation of water leakage areas in large-scale underground infrastructure with Tunnel Monorail Imaging
Core Problem: Manual detection and documentation of water leakage in large-scale underground infrastructure (tunnels) is inefficient and inaccurate, posing risks to asset management and structural integrity.
Key Innovation: A computer vision-based method using a robot-mounted 360-degree camera and an ensemble deep learning strategy is developed to automatically demarcate and map tunnel leakage areas, enhancing monitoring efficiency and accuracy for large-scale underground infrastructure.
59. Experimental investigation of high-temperature fire resistance properties of corrugated steel-concrete composite structures
Core Problem: Fire poses a severe threat to the safety of corrugated steel-lined tunnels, and a detailed analysis of their damage and failure mechanisms under fire conditions is needed.
Key Innovation: Full-scale fire tests on corrugated steel-concrete composite structures under thermo-mechanical coupling effects are conducted to investigate their mechanical response and damage mechanism, validating the fire resistance performance and providing critical experimental evidence for tunnel lining design.
60. Study on mechanical properties of self-prestressed prefabricated lining pipe for trenchless rehabilitation of large diameter underground pipelines
Core Problem: Structurally compromised large-diameter underground pipelines require effective trenchless rehabilitation technologies that can provide sufficient mechanical performance and long-term durability.
Key Innovation: An innovative trenchless rehabilitation technology using self-prestressed prefabricated ultra-high-performance concrete (UHPC) lining pipes reinforced with iron-based shape memory alloy (Fe-SMA) rebars is introduced, demonstrating significantly higher ultimate strength and cracking load compared to ordinary steel rebars.
61. Shear performance of prefabricated utility tunnel joints considering the effect of connectors
Core Problem: Joints in prefabricated utility tunnels are vulnerable to cracking, leakage, and misalignment, seriously compromising operational safety and reliability.
Key Innovation: The influence of connectors on the shear failure characteristics and mechanical behavior of prefabricated utility tunnel joints is investigated through scale model tests and 3D numerical simulation, clarifying internal force distribution and failure mechanisms, and recommending optimization for enhanced joint stability and bearing capacity.
62. Data strategies for automatic calibration of soil constitutive models
Core Problem: The reliability of automatic calibration (AC) tools for soil constitutive models depends on the quality and composition of experimental datasets, and optimal data strategies (weighting, test combinations, minimum tests) are not well-defined.
Key Innovation: Examined dataset composition and weighting strategies for automatic calibration of the SANISAND-F model using extensive element tests, finding that incorporating at least two distinct test types significantly enhances calibration stability and that as few as 8 appropriately selected tests can yield stable optimization results, providing practical guidance for experimental design.
63. Probabilistic NDVI Forecasting from Sparse Satellite Time Series and Weather Covariates
Core Problem: Accurate short-term forecasting of vegetation dynamics (NDVI) is challenging due to sparse and irregular satellite sampling (caused by cloud coverage) and heterogeneous climatic conditions, hindering data-driven decision support in precision agriculture.
Key Innovation: Proposes a probabilistic forecasting framework using a transformer-based architecture that integrates historical NDVI observations with both historical and future meteorological covariates. It introduces a temporal-distance weighted quantile loss and incorporates cumulative/extreme-weather feature engineering to address irregular revisit patterns and capture delayed meteorological effects, outperforming baselines.
64. Parallel Complex Diffusion for Scalable Time Series Generation
Core Problem: Traditional temporal diffusion models for time series generation struggle with modeling long-range dependencies due to local entanglement and high computational cost (O(L^2)) of attention mechanisms, limiting scalability and efficiency.
Key Innovation: Introduces PaCoDi (Parallel Complex Diffusion), a spectral-native architecture that decouples generative modeling in the frequency domain, converting locally coupled temporal signals into globally decorrelated spectral components. It proves a Quadrature Forward Diffusion and Conditional Reverse Factorization theorem, uses a Mean Field Theory approximation, generalizes to continuous-time Frequency SDEs, and exploits Hermitian Symmetry for 50% FLOP reduction, outperforming baselines in generation quality and inference speed.
65. VidEoMT: Your ViT is Secretly Also a Video Segmentation Model
Core Problem: Existing online video segmentation models combine per-frame segmenters with complex specialized tracking modules, introducing significant architectural complexity and computational overhead.
Key Innovation: Proposes VidEoMT, a simple encoder-only video segmentation model that eliminates dedicated tracking modules by introducing a lightweight query propagation mechanism and a query fusion strategy, achieving competitive accuracy at 5x-10x faster speeds.
66. On the Evaluation Protocol of Gesture Recognition for UAV-based Rescue Operation based on Deep Learning: A Subject-Independence Perspective
Core Problem: Existing deep learning-based gesture recognition evaluations for UAV-based rescue operations suffer from data leakage due to improper train-test splits, leading to inflated accuracy metrics that do not reflect generalization to unseen individuals.
Key Innovation: A methodological analysis demonstrating the flaws in a specific evaluation protocol, highlighting the critical importance of subject-independent data partitioning for reliable gesture recognition in applications like UAV-human interaction for rescue operations.
67. COMBA: Cross Batch Aggregation for Learning Large Graphs with Context Gating State Space Models
Core Problem: Advancing State Space Models (SSMs) to effectively learn from large graph-structured data is challenging due to SSMs being sequence models and the computational cost of converting graphs to sequences.
Key Innovation: Proposed COMBA, a framework using graph context gating and cross-batch aggregation to scale SSMs for large graph learning, demonstrating significant performance gains and theoretical guarantees for lower error.
68. PHAST: Port-Hamiltonian Architecture for Structured Temporal Dynamics Forecasting
Core Problem: Reliably forecasting the long-horizon dynamics of dissipative physical systems from only partial (position-only) observations, while ensuring stability and recovering physically meaningful parameters.
Key Innovation: Introduction of PHAST (Port-Hamiltonian Architecture for Structured Temporal dynamics), which leverages the port-Hamiltonian framework to explicitly decompose the Hamiltonian, guaranteeing stable long-horizon forecasts and enabling physically meaningful parameter recovery across diverse physical systems.
69. RoEL: Robust Event-based 3D Line Reconstruction
Core Problem: Event cameras in motion produce sparse line data, which can lead to drastic deterioration with minor estimation errors, making robust 3D line reconstruction and pose refinement challenging, especially with unpredictable noise characteristics.
Key Innovation: RoEL, a method that stably extracts tracks of varying line appearances by observing multiple representations from various time slices of events, and proposes geometric cost functions to refine 3D line maps and camera poses, demonstrating significant performance boost in event-based mapping and pose refinement.
70. Quantum-enhanced satellite image classification
Core Problem: Enhancing the accuracy of multi-class image classification for space applications, particularly in satellite imaging and remote sensing, beyond what robust classical methods can achieve.
Key Innovation: Demonstrates a quantum-classical hybrid approach using quantum feature extraction (based on many-body spin Hamiltonians) combined with classical processing, achieving a 2-3% absolute accuracy improvement over strong classical baselines (e.g., ResNet50) in satellite image classification.
71. Scaling Laws in Patchification: An Image Is Worth 50,176 Tokens And More
Core Problem: Understanding the information loss caused by patchification-based compressive encoding in Vision Transformers and its effect on visual understanding.
Key Innovation: Discovery of a 'scaling law in patchification' where models consistently benefit from decreased patch sizes down to 1x1 pixel tokenization across various tasks and architectures, leading to improved predictive performance and insights for non-compressive vision models.
72. Assimilative Causal Inference
Core Problem: Existing causal inference methods struggle to capture instantaneous, time-evolving causal relationships in complex, high-dimensional systems, particularly in identifying dynamic causal interactions without observing candidate causes or accommodating short datasets.
Key Innovation: Develops Assimilative Causal Inference (ACI), a methodological framework leveraging Bayesian data assimilation to trace causes backward from observed effects, uniquely identifying dynamic causal interactions, accommodating short datasets, and providing online tracking of causal roles and a criterion for causal influence range, demonstrated on complex dynamical systems with intermittency and extreme events.
73. Uncertainty Estimation by Flexible Evidential Deep Learning
Core Problem: Existing Evidential Deep Learning (EDL) methods for uncertainty quantification (UQ) are limited by the restrictive assumption of Dirichlet-distributed class probabilities, which reduces their robustness and generalization in complex or unforeseen scenarios.
Key Innovation: Proposes Flexible Evidential Deep Learning (F-EDL) which extends EDL by predicting a flexible Dirichlet distribution over class probabilities, providing a more expressive and adaptive representation of uncertainty. This significantly enhances UQ generalization and reliability, demonstrating state-of-the-art performance across diverse evaluation settings.
74. MIST: Mutual Information Estimation Via Supervised Training
Core Problem: Traditional mutual information (MI) estimators often trade universal theoretical guarantees for flexibility and efficiency, while existing neural baselines are slow for inference and lack robust uncertainty quantification.
Key Innovation: Proposes MIST, a fully data-driven approach that parameterizes MI estimation with a neural network trained end-to-end on a large meta-dataset of synthetic joint distributions with known ground-truth MI. MIST outperforms classical baselines in speed and accuracy, provides well-calibrated quantile-based uncertainty intervals, and offers a trainable, differentiable estimator embeddable in larger learning pipelines.
75. Dynamic response analysis of mooring cables for floating offshore wind turbines under ocean currents using a perturbation-based finite difference method
Core Problem: Efficiently and accurately analyzing the dynamic response of mooring cables for floating offshore wind turbines under complex ocean current conditions, which is crucial for overall system performance.
Key Innovation: An efficient dynamic model for mooring cables that incorporates hydrodynamic effects using a perturbation-based finite difference method (Keller-box scheme) and the Bisection method, achieving substantial improvements in computational efficiency and accuracy for analyzing dynamic response under wave and current loading.
76. The study of foam stability during EPB shield tunnelling based on surface tension and a new half-life time measurement method
Core Problem: Foam stability is critical for effective soil conditioning in EPB shield tunnelling, but existing half-life time (hlt) measurement methods have limitations that hinder accurate assessment.
Key Innovation: A new method for measuring foam half-life time (hlt) is proposed, relying on monitoring foam drainage at a fixed foam column height, which clarifies the effects of foam expansion ratio and foaming agent concentration on stability, linking them to surface tension and drainage patterns.
77. Upper-bound solutions for capacity of tracked vehicles on clay under fully three-dimensional loading
Core Problem: Lack of rigorous analytical foundations for predicting the ultimate bearing capacity of tracked vehicles on clay under fully three-dimensional loading conditions, extending beyond conventional coplanar limitations.
Key Innovation: Developed novel upper-bound solutions for predicting ultimate bearing capacity by integrating the superposition of rotation-admissible Green mechanisms with two orthogonal variable cross-section planar velocity fields, establishing the first rigorous analytical foundation for six-degree-of-freedom bearing capacity analysis of tracked vehicles.
78. Alternative approach for the calibration of reliability-based regional resistance factors for axially loaded drilled shafts
Core Problem: National code resistance factors for drilled shafts may not accurately reflect regional design uncertainty and often fail to achieve desired reliability levels, while regional calibration is hindered by limited quality load test data.
Key Innovation: Proposed and demonstrated a segmental procedure based on strain gauge data to overcome the challenge of limited load test data in statistically characterizing side resistance uncertainties, enabling the calibration of more accurate reliability-based regional resistance factors for axially loaded drilled shafts.
79. Simple method for the numerical simulation of the elastic impulse response of a pile head
Core Problem: The simple pile load test, being dynamic, requires a straightforward simulation method to extrapolate results to the low-frequency range for static values, making it practical for field applications.
Key Innovation: Proposed a simple calculation method for the elastic impulse response at a pile head by extending the Winkler model with first-order approximations for load distribution and Green's function, successfully replicating results from physical tests and finite element analyses to enable practical extrapolation of dynamic pile load test results to static values.
80. Oxygen Isotopes in Tree Rings Track Neotropical Climate Dynamics
Core Problem: Central America faces increasing risks from climate variability and extreme weather events, but limited observational records and model biases constrain the understanding of ocean-atmosphere dynamics influencing regional precipitation over long timescales.
Key Innovation: Development of a new multi-decadal tree-ring δ18O record from Abies guatemalensis in Guatemala and Honduras, demonstrating its tight coupling to boreal summer rainfall and neotropical ocean-atmosphere dynamics, providing a precisely dated, high-resolution proxy for multi-century hydroclimate reconstructions.
81. Detectability of Phytoplankton Biomass Extremes Using Simulated Satellite Chlorophyll Observations
Core Problem: Satellite chlorophyll observations, often used to study extreme phytoplankton events, provide only a surface view, are affected by physiological variability, and have missing data, leading to poor understanding of the global occurrence, drivers, and consequences of these events.
Key Innovation: Used an Earth system model with a satellite chlorophyll simulator to quantify the detectability of vertically integrated phytoplankton biomass extremes, revealing that only a small fraction (10-19%) are detected and that many detected chlorophyll extremes do not correspond to true biomass extremes.
82. The Impact of OCO‐2 Seasonally Dependent Sampling on Carbon Flux Estimation in the Northern Tropical Africa
Core Problem: The large annual carbon source over northern tropical Africa inferred from satellite CO2 remains debated, with a hypothesis that seasonally dependent sampling by OCO-2 might lead to overestimated annual fluxes due to uneven data collection during growing vs. non-growing seasons.
Key Innovation: Used observing system simulation experiments with OCO-2 sampling to demonstrate that seasonally dependent sampling, especially when prior flux seasonal cycles differ from truth, can lead to overestimated annual carbon fluxes, highlighting the need for improved prior estimates and expanded observational coverage.
83. Attributing Long‐Term Trends in Marine Low Cloud Morphologies to Aerosols and Large‐Scale Meteorology With Deep Learning
Core Problem: The response of marine low-cloud mesoscale morphologies to climate change and emission reductions, and the attribution of long-term trends in these morphologies to aerosols and large-scale meteorology, remain poorly understood.
Key Innovation: Developed a deep learning model (UMorNet) to predict instantaneous cloud morphologies from meteorology and aerosol proxies, achieving good accuracy and capturing spatial patterns of climatology and long-term trends, identifying key drivers like cloud droplet number concentration, marine cold-air outbreak index, SST, and inversion strength.
84. Soil Structure and Mixing Controls on Water‐Rock Contact: Implications for Enhanced Weathering
Core Problem: Quantifying weathering rates in soils for enhanced weathering (EW) is challenging because most continuum-scale EW models do not adequately account for the wet fraction of rock powder surface area that is actively weathering, leading to overestimations.
Key Innovation: Developed a mechanistic, computationally efficient framework based on soil physics to derive a scaling factor that quantifies the wet fraction of rock powder surface area as a function of soil moisture and mixing degree, improving continuum-scale EW models and explaining discrepancies in observed weathering rates.
85. Investigating Target Class Influence on Neural Network Compressibility for Energy-Autonomous Avian Monitoring
Core Problem: Traditional avian monitoring is costly and inefficient; existing machine learning solutions for passive acoustic monitoring require substantial computational resources, making them unsuitable for energy-autonomous edge devices in the field.
Key Innovation: Developing a method for avian monitoring on microcontrollers (MCUs), training and compressing models for various numbers of target classes, and demonstrating significant compression rates with minimal performance loss, enabling the deployment of energy-autonomous devices for wildlife monitoring.
86. Multi-material Multi-physics Topology Optimization with Physics-informed Gaussian Process Priors
Core Problem: Existing machine learning-based topology optimization (TO) methods struggle with high computational cost, spectral bias, and handling complex multi-material, multi-physics problems, especially when objective or constraint functions are not self-adjoint.
Key Innovation: Proposing a framework based on physics-informed Gaussian processes (PIGPs) for multi-material, multi-physics topology optimization. This framework represents primary, adjoint, and design variables with independent GP priors whose mean functions are neural networks, enabling simultaneous solution of coupled multi-physics and design problems, generating super-resolution topologies with sharp interfaces and physically interpretable material distributions.
87. Neural Prior Estimation: Learning Class Priors from Latent Representations
Core Problem: Class imbalance in deep neural networks leads to systematic bias and skewed effective class priors, hindering accurate prediction, especially for underrepresented classes.
Key Innovation: Introduction of Neural Prior Estimator (NPE), a framework that learns feature-conditioned log-prior estimates from latent representations, providing a theoretically grounded adaptive signal for bias-aware prediction without explicit class counts or distribution-specific hyperparameters.
88. JAX-Privacy: A library for differentially private machine learning
Core Problem: The complexity and difficulty of deploying robust and performant mechanisms for differentially private machine learning, hindering both researchers and practitioners.
Key Innovation: Introduction of JAX-Privacy, a library designed to simplify the deployment of differentially private machine learning mechanisms by providing verified, modular primitives for critical components, catering to both researchers and practitioners.
89. Financial time series augmentation using transformer based GAN architecture
Core Problem: The scarcity and dynamic nature of financial time series data limit the effective training and generalization of advanced deep learning models for forecasting, leading to sub-optimal predictive accuracy.
Key Innovation: Demonstration that Generative Adversarial Networks (GANs), specifically a transformer-based GAN (TTS-GAN), can effectively augment scarce financial time series data, significantly improving the forecasting accuracy of LSTM models. Also proposes a novel time series specific quality metric.
90. MantisV2: Closing the Zero-Shot Gap in Time Series Classification with Synthetic Data and Test-Time Strategies
Core Problem: A substantial performance gap exists between frozen and fine-tuned encoders in time series foundation models, limiting their zero-shot feature extraction capabilities for diverse downstream tasks.
Key Innovation: Introduction of MantisV2 and Mantis+, which significantly strengthen zero-shot feature extraction for time series by pre-training on synthetic data, refining the encoder architecture, and proposing enhanced test-time methodologies (leveraging intermediate-layer representations, refined output-token aggregation, self-ensembling, cross-model embedding fusion), achieving state-of-the-art zero-shot performance.
91. Learning Compact Video Representations for Efficient Long-form Video Understanding in Large Multimodal Models
Core Problem: Analyzing long-form videos with contemporary state-of-the-art models is challenging due to the inherent redundancy of video sequences, leading to difficulties in efficiently incorporating many frames within memory constraints and extracting discriminative information.
Key Innovation: Introduction of a novel end-to-end schema for long-form video understanding, integrating an information-density-based adaptive video sampler (AVS) and an autoencoder-based spatiotemporal video compressor (SVC) with an MLLM. This system adaptively captures essential information and achieves high compression rates while preserving crucial discriminative information.
92. Understanding the Fine-Grained Knowledge Capabilities of Vision-Language Models
Core Problem: Despite substantial progress in general visual question answering, Vision-Language Models (VLMs) still trail behind in traditional image classification benchmarks that test fine-grained visual knowledge, indicating a disconnect in their capabilities.
Key Innovation: A series of ablation experiments identifying factors contributing to the disconnect between fine-grained knowledge and other vision benchmarks in VLMs. Findings show that better LLMs improve all scores equally, while better vision encoders disproportionately improve fine-grained classification, and the pretraining stage (especially unfrozen LLM weights) is vital for fine-grained performance.
93. Breaking the Correlation Plateau: On the Optimization and Capacity Limits of Attention-Based Regressors
Core Problem: Attention-based regression models often exhibit a "PCC plateau" during training, where the Pearson correlation coefficient stops improving early, even as MSE decreases, due to conflicts between magnitude and shape matching and capacity limits of convex aggregators.
Key Innovation: Provided the first rigorous theoretical analysis of the PCC plateau, identifying optimization conflicts and capacity limits, and proposed Extrapolative Correlation Attention (ECA) to break this plateau, achieving significant correlation improvements without compromising MSE.
94. Optimizing Graph Causal Classification Models: Estimating Causal Effects and Addressing Confounders
Core Problem: Traditional graph machine learning methods, including GNNs, rely on correlations and are sensitive to spurious patterns and distribution changes, making them less robust for understanding true cause-effect relationships and handling confounders in complex systems.
Key Innovation: Proposes CCAGNN, a Confounder-Aware causal GNN framework that integrates causal reasoning into graph learning, supporting counterfactual reasoning and providing more reliable and robust predictions by isolating true causal factors and adjusting for confounders.
95. Image Quality Assessment: Exploring Quality Awareness via Memory-driven Distortion Patterns Matching
Core Problem: The limitation of existing full-reference image quality assessment (FR-IQA) methods, which rely heavily on high-quality reference images, hindering their real-world application where such references are often unavailable.
Key Innovation: A memory-driven quality-aware framework (MQAF) that establishes a memory bank of distortion patterns and dynamically switches between dual-mode (full-reference and no-reference) quality assessment strategies, significantly reducing reliance on reference images and outperforming state-of-the-art approaches.
96. NIMMGen: Learning Neural-Integrated Mechanistic Digital Twins with LLMs
Core Problem: Existing LLM-based agentic frameworks for constructing mechanistic models oversimplify real-world conditions, leading to unclear reliability in practice, especially with partial observations and diverse task objectives.
Key Innovation: Introduces the NIMM evaluation framework for realistic assessment of LLM-generated mechanistic models and NIMMgen, an agentic framework that enhances code correctness and practical validity through iterative refinement, demonstrating strong performance across diverse scientific domains and supporting counterfactual intervention simulation.
97. Stable Long-Horizon Spatiotemporal Prediction on Meshes Using Latent Multiscale Recurrent Graph Neural Networks
Core Problem: Accurate, stable, and generalizable long-horizon prediction of spatiotemporal fields on complex geometries, where high-fidelity simulations are costly and machine learning methods struggle with stability and generalization over extended time steps.
Key Innovation: A deep learning framework for predicting full spatiotemporal histories directly on meshes, using a temporal multiscale architecture with coupled latent recurrent graph neural networks and a variational graph autoencoder, achieving accurate and temporally stable long-horizon predictions across diverse geometries.
98. Parameter-Efficient Domain Adaptation of Physics-Informed Self-Attention based GNNs for AC Power Flow Prediction
Core Problem: Accurate AC power flow (AC-PF) prediction is critical for power grids, but models trained on medium-voltage (MV) grids suffer performance degradation when deployed on high-voltage (HV) networks (domain shift), and existing full fine-tuning methods are costly and lack control over stability-plasticity trade-off.
Key Innovation: Parameter-efficient domain adaptation for physics-informed self-attention based GNNs using LoRA applied to attention projections with selective unfreezing of the prediction head, encouraging Kirchhoff-consistent behavior via a physics-based loss, achieving near-full fine-tuning accuracy with significantly fewer trainable parameters and controllable adaptation capacity.
99. Self-Aware Object Detection via Degradation Manifolds
Core Problem: Object detectors fail silently under various image degradations (blur, noise, adverse weather), lacking the ability to assess if input is within their nominal operating regime, which is critical for safety-critical applications.
Key Innovation: Introduces a degradation-aware self-awareness framework based on degradation manifolds, structuring the detector's feature space by image degradation. It uses multi-layer contrastive learning to pull similar degradations together and push different ones apart, defining a pristine prototype for self-awareness as geometric deviation.
100. Going Down Memory Lane: Scaling Tokens for Video Stream Understanding with Dynamic KV-Cache Memory
Core Problem: Existing state-of-the-art methods for streaming video understanding, which rely on key-value caching, use a limited number of tokens per frame, leading to the loss of fine-grained visual details and a bias towards later frames in retrieval due to increasing query-frame similarity scores over time.
Key Innovation: Proposes MemStream, a method for scaling the token budget for more granular spatiotemporal understanding. It introduces an adaptive selection strategy to reduce token redundancy while preserving local spatiotemporal information and a training-free retrieval mixture-of-experts to better identify relevant frames, significantly improving performance on video question answering benchmarks.
101. IRPAPERS: A Visual Document Benchmark for Scientific Retrieval and Question Answering
Core Problem: Visual document processing remains underexplored compared to text/relational data, and it is unclear how recent image-based multimodal foundation models compare to established text-based methods for scientific document retrieval and question answering.
Key Innovation: Introduces IRPAPERS, a benchmark of scientific papers with both image and OCR transcriptions, demonstrating that multimodal hybrid search (combining image and text retrieval) outperforms unimodal approaches and identifying complementary limitations, providing a resource for advancing scientific document AI.
102. Drift Estimation for Stochastic Differential Equations with Denoising Diffusion Models
Core Problem: Estimating time-homogeneous drift functions in multivariate stochastic differential equations (SDEs) from observed trajectories, especially in higher dimensions, remains a challenge for classical methods.
Key Innovation: A novel approach formulates drift estimation as a denoising problem and proposes an estimator derived as a by-product of training a conditional diffusion model. This method effectively estimates drift functions in SDEs, matching classical methods in low dimensions and remaining competitive in higher dimensions, while also enabling dynamic simulation of new trajectories.
103. Learning from Biased and Costly Data Sources: Minimax-optimal Data Collection under a Budget
Core Problem: Naive data collection strategies and standard estimators are suboptimal when gathering data from multiple biased and costly sources to estimate population means under a fixed budget.
Key Innovation: Develops a sampling plan that maximizes effective sample size (based on χ²-divergence) paired with a post-stratification estimator, achieving budgeted minimax optimal risk for estimating population and group-conditional means, and extends to prediction problems.
104. SUNLayer: Stable denoising with generative networks
Core Problem: Providing a rigorous and simplified theoretical analysis of generative models for image denoising, particularly identifying conditions for stable denoising under local optimization.
Key Innovation: Introduces SUNLayer, a theoretical framework based on spherical harmonics, which identifies explicit conditions on activation functions that guarantee stable denoising with generative networks under local optimization, validated by numerical experiments.
105. GIFT: A Framework Towards Global Interpretable Faithful Textual Explanations of Vision Classifiers
Core Problem: Existing explainability approaches for deep vision models suffer from limited faithfulness, local scope, or ambiguous semantics, hindering trustworthy deployment.
Key Innovation: GIFT, a post-hoc framework that generates local visual counterfactuals, translates them into natural language, aggregates them into global hypotheses using LLMs, and verifies them through image-based interventions, providing causally grounded textual explanations for vision models.
106. Analyzing the Training Dynamics of Image Restoration Transformers: A Revisit to Layer Normalization
Core Problem: Conventional Layer Normalization (LN) in Image Restoration (IR) Transformers causes feature magnitudes to diverge and collapses channel-wise entropy due to misalignments with IR tasks (per-token normalization disrupting spatial correlations, input-independent scaling discarding input-specific statistics).
Key Innovation: Proposes Image Restoration Transformer Tailored Layer Normalization (i-LN), a simple drop-in replacement that normalizes features holistically and adaptively rescales them per input, leading to improved training dynamics and performance in IR tasks.
107. Beyond Simple Graphs: Neural Multi-Objective Routing on Multigraphs
Core Problem: Existing learning-based routing methods are unsuitable for multi-objective routing on multigraphs, which are prevalent in real-world scenarios with multiple distinct attributes between node pairs.
Key Innovation: Proposes two graph neural network-based methods for multi-objective routing on multigraphs: one directly operates on the multigraph, and a more scalable one simplifies the multigraph via learned pruning before routing, demonstrating competitive performance.
108. Comparative Analysis of Wave Scattering Numerical Modeling Using the Boundary Element Method and Physics-Informed Neural Networks
Core Problem: Evaluating the performance of Boundary Element Method (BEM) and Physics-Informed Neural Networks (PINNs) for solving wave scattering problems (Helmholtz equation) in terms of accuracy and computational efficiency.
Key Innovation: A direct comparative analysis demonstrating that BEM is significantly faster for assembly and solution, while a trained PINN can achieve faster evaluation times at interior points, establishing a procedure for comparing these methods for wave propagation problems.
109. Study of Training Dynamics for Memory-Constrained Fine-Tuning
Core Problem: Training large deep neural networks is memory-intensive, posing significant challenges for fine-tuning and deployment in environments with strict resource constraints.
Key Innovation: Introduces TraDy, a novel transfer learning scheme that leverages architecture-dependent layer importance and dynamic stochastic channel selection to achieve state-of-the-art performance in memory-constrained fine-tuning. It significantly reduces activation sparsity (up to 99%), weight derivative sparsity (95%), and FLOPs for weight derivative computation (97%).
110. Phase-space entropy at acquisition reflects downstream learnability
Core Problem: The lack of a general, modality-agnostic way to quantify how data acquisition itself preserves or destroys information crucial for downstream machine learning tasks, especially under aggressive undersampling.
Key Innovation: Proposal of $\Delta S_{\mathcal B}$, an acquisition-level scalar based on instrument-resolved phase space, which directly quantifies how acquisition mixes or removes joint space-frequency structure and consistently predicts downstream reconstruction/recognition difficulty across modalities without training.
111. UrbanGS: A Scalable and Efficient Architecture for Geometrically Accurate Large-Scene Reconstruction
Core Problem: Existing 3D Gaussian Splatting (3DGS) methods struggle with geometric consistency, memory efficiency, and computational scalability when applied to large-scale urban environments.
Key Innovation: UrbanGS, a scalable reconstruction framework that uses a Depth-Consistent D-Normal Regularization module with external depth supervision for enhanced geometric accuracy, and a Spatially Adaptive Gaussian Pruning (SAGP) strategy for improved scalability and memory efficiency in large-scale urban scene reconstruction.
112. Multi-View Wireless Sensing via Conditional Generative Learning: Framework and Model Design
Core Problem: Achieving high-precision target sensing using multi-view channel state information (CSI) in wireless networks requires effectively fusing latent target features and reconstructing complex target properties while incorporating physical knowledge.
Key Innovation: Gen-MV, a conditional generative multi-view sensing framework, employs a bipartite neural network (encoder for multi-view CSI fusion, conditional diffusion model for target reconstruction) and incorporates physical knowledge and spatial positional embedding to achieve excellent flexibility and significant performance improvement in reconstructing target shape and electromagnetic properties.
113. A joint optimization approach to identifying sparse dynamics using least squares kernel collocation
Core Problem: Accurately learning systems of ordinary differential equations (ODEs) from scarce, partial, and noisy observations, improving upon existing methods in terms of accuracy, sample efficiency, and noise robustness.
Key Innovation: Develops an all-at-once modeling framework combining sparse recovery for ODEs over a function library with RKHS techniques for state estimation and ODE discretization, demonstrating significant gains over existing algorithms.
114. Anomaly diagnosis of dynamic cable state based on LSTM-SAE
Core Problem: Detecting early micro-damage to dynamic cables, critical components in offshore wind projects, is challenging, leading to significant economic implications if failures occur.
Key Innovation: A dynamic cable structural anomaly diagnosis method and framework integrating Long Short-Term Memory network (LSTM) and Stacked Auto-Encoder (SAE). It uses an extremum prediction algorithm for dynamic threshold determination and constructs a tension prediction model to diagnose cable status based on prediction errors, capable of detecting damage above 4%.
115. A theoretical approach for axial compressive analysis of helical piles supporting offshore jacket structures
Core Problem: A dearth of research on the compressive capacity of helical piles associated with settlement, especially for multi-helix piles, which impedes the advancement of response-based design methodologies for offshore jacket structures.
Key Innovation: A comprehensive unified theoretical framework for analyzing the axial compressive response of multi-helix piles, integrating individual bearing and cylindrical shear modes using a multi-fictitious soil pile model and a semi-analytical solution via the transfer matrix method.
116. A 25 km Daily Gridded Dataset of Meteorological Variables and High-Impact Weather Events for New-type Power Systems in China (1980–2016)
Core Problem: The 'weather dependency' of new-type power systems and the lack of a high-quality gridded dataset combining meteorological variables and high-impact weather events relevant to their operational security in China.
Key Innovation: Development of the China New-type Power Systems Meteorological (CNPS-Met) dataset, a daily, 25 km resolution gridded dataset (1980–2016) for China, including eight meteorological variables and eleven high-impact weather events, generated using a spatially adaptive optimal interpolation scheme, demonstrating superior performance in meteorological estimation.
117. Tectono-stratigraphic evolution of the Upper Kaghan Valley, Northwestern Himalayas: insights from integrated field and remote-sensing analysis
Core Problem: Earlier studies lacked a unified tectono-stratigraphic analysis incorporating all necessary geological, structural, and remote-sensing information for the Upper Kaghan Valley.
Key Innovation: Presents a unified tectono-stratigraphic column for the Upper Kaghan Valley using integrated field mapping, structural analysis, and Sentinel-1 SAR, constraining deformation and uplift sequences.
118. Comparison of edge enhancement methods for potential field data: applications to a synthetic model and the Laxmi Basin, Arabian Sea
Core Problem: Accurately delineating subsurface source boundaries from gravity and magnetic data using various edge-enhancement techniques, and evaluating their performance and robustness.
Key Innovation: Systematically compares common edge-enhancement techniques, finding that Wavelet Space Entropy (WSE) provides superior noise robustness and reliably delineates geological boundaries in both synthetic and real-world data.
119. Novel sealing concepts for landfills: mechanism and prediction of permeabilities of bentonite-sand mixtures
Core Problem: Landfills pose a significant risk of groundwater contamination due to leachate, requiring enhanced impermeable systems beyond traditional bentonite-sand composites.
Key Innovation: Investigated an alternative impermeable system using quarry stone chips and bentonite, demonstrating a significant reduction in permeability with increased bentonite content. Introduced a novel evaluation method for determining the permeability coefficient of mixed soil.
120. Prediction of lithium content in typical mountainous clay in Xinjiang, China using fractional derivatives and feature extraction
Core Problem: The need for a fast, non-destructive method to estimate lithium (Li) content in mountainous clay soils using hyperspectral techniques, as Li is an emerging environmental pollutant.
Key Innovation: Developed a new method combining fractional order derivatives (FOD) and Random Forest (RF) for rapid and non-destructive estimation of soil lithium content, demonstrating improved model performance and providing a theoretical basis for hyperspectral exploration of Li resources.
121. Exploring patterns and impacts of farmland and construction land upslope in China: based on an integrated slope spectrum and sensitivity model
Core Problem: Relying solely on the slope spectrum is too broad to understand the underlying formations and impacts of farmland and construction land development upslope in China, hindering effective regional management strategies.
Key Innovation: An integrated approach combining the slope spectrum with a novel slope sensitivity coefficient (SSC) derived from land use transfer matrices to explore and classify patterns of farmland and construction land upslope (passive, active, saturation, avoidance), revealing their distinct impacts on food security, ecological protection, and city livability, providing a reference for sustainable land use strategies.
122. Assessment of the potential of sub-diurnal in situ C-band radar data for monitoring maize crop in semi-arid regions
Core Problem: Investigating the potential of sub-diurnal C-band radar data for monitoring maize crop, particularly its response to plant water stress and physiological activity.
Key Innovation: Demonstrates that temporal coherence exhibits a daily cycle linked to physiological activity, and daily mean backscattering coefficient correlates with evapotranspiration and evaporative fraction, showing potential for detecting water stress and disentangling physiological from structural drivers.
123. Enhanced remote sensing of surface water Chlorophyll-a: Coupling dynamic algae vertical movement modeling with multi-spectral satellite images
Core Problem: Current remote sensing approaches for water quality monitoring face limitations in aligning satellite observations with in-situ measurements due to dynamic vertical behavior of algae and temporal constraints of satellite overpasses.
Key Innovation: Proposes an Algal Behavior Function (ABF) to model diurnal vertical migration of algae, enabling temporal adjustment of in-situ measurements to generate refined field-to-satellite matchups, which improved Chl-a prediction accuracy by 5.8%-18.0% and showed moderate geographic transferability.
124. Estimation of global riverine total phosphorus concentration based on multi-source data and stacked ensemble learning
Core Problem: Quantifying riverine total phosphorus (TP) concentration at the global scale using remote sensing is challenging because TP is not optically active and its variability is regulated by complex processes.
Key Innovation: Integrates satellite-derived reflectance with environmental predictors and employs stacked ensemble learning (R2 = 0.80) to estimate global riverine TP concentration, demonstrating improved stability and transferability across heterogeneous river systems.
125. Calculation and prediction of CO2 concentrations inside a ventilation gallery of Madrid Calle 30 urban tunnels
Core Problem: High CO2 concentrations in urban tunnels lead to concrete carbonation, reducing infrastructure lifespan, and there is a need for reliable methods to calculate and predict these concentrations for carbonation prediction models.
Key Innovation: Developed a methodology to calculate and forecast CO2 concentrations inside urban tunnels based on traffic intensity data and the composition of the circulating fleet. This methodology can be adapted for concrete carbonation analysis in other urban tunnels, aiding in infrastructure lifespan management.
126. Remote control of North China autumn rainfall by Tibetan Plateau soil conditions
Core Problem: The drivers of autumn rainfall in North China, particularly extreme rainfall events, remain poorly understood, impacting agriculture and water resources.
Key Innovation: Revealed that Tibetan Plateau subsurface soil temperature and moisture anomalies significantly modulate North China's autumn total and extreme rainfall, explaining over a quarter of East Asian autumn precipitation variability, and offering predictive potential for seasonal and weather forecasts for extreme weather management.