Curated, top-journal papers on foundation models and adjacent AI paradigms for Earth and environmental applications, sorted by publication year (newest first). This list is updated on a rolling basis.

Year	Journal	Title	Key Summary	Link
2026	Communications Earth & Environment	On the foundations of Earth foundation models	This perspective proposes 11 key features for ideal Earth foundation models, including geolocation embedding, scale awareness, multisensor integration, carbon-efficient operation, and physical consistency. It discusses current limitations, evaluation strategies, and future directions such as energy-efficient adaptation, adversarial robustness, and interpretability for Earth observation, weather, climate, and sustainability applications.	View Paper
2025	Nature	A foundation model for the Earth system	This paper introduces Aurora, a large-scale foundation model pretrained on more than one million hours of geophysical data (ERA5, CAMS, MERRA-2, and others). Aurora outperforms several operational systems across air quality forecasting, wave prediction, tropical cyclone tracking, and 10-day high-resolution weather forecasting, while requiring much lower computational cost. The results show scalable, general-purpose representations that can be fine-tuned across Earth system tasks.	View Paper
2025	Nature	End-to-end data-driven weather prediction	This study presents Aardvark Weather, an end-to-end machine learning system that replaces conventional numerical weather prediction pipelines with a single model trained directly on raw observations from satellites, buoys, radiosondes, and other sources. It delivers global to local forecasts up to 10 days ahead with strong accuracy and speed, and demonstrates foundation-model-style direct data-to-forecast learning for practical use in agriculture, energy, and disaster management.	View Paper
2025	Nature Machine Intelligence	Towards responsible geospatial foundation models	This editorial discusses the development of geospatial foundation models based on Earth observation data for ecology, urban computing, and remote sensing. It reviews 58 remote-sensing vision foundation models from 2021 to 2024, highlights multimodal pretraining strategies, and emphasizes sustainability, resource efficiency, and privacy risks when scaling high-resolution geospatial AI.	View Paper
2024	Nature	Neural general circulation models for weather and climate	This work introduces NeuralGCM, a hybrid general circulation model that combines a differentiable dynamical core with machine learning parameterizations for unresolved processes such as clouds and radiation. Trained on ERA5, it matches or exceeds strong baselines in 1 to 15 day weather prediction and reproduces realistic multi-decadal climate statistics, showing a scalable pathway for physically grounded ML in Earth and environmental prediction.	View Paper

arXiv-Related Foundation Model Papers

Year	Title	Key Summary	Link
2026	A Genealogy of Foundation Models in Remote Sensing	This survey reviews the evolution of remote-sensing foundation models from single-sensor to multisensor designs, covering pretraining strategies, adaptation methods, and benchmark trends. It highlights constraints such as high computational demands and data imbalance, and outlines future directions for robust geospatial representations.	arXiv:2504.17177
2025	First On-Orbit Demonstration of a Geospatial Foundation Model	This paper reports the first on-orbit deployment of a geospatial foundation model (IMAGIN-e on the ISS) and evaluates compression and domain-adaptation strategies across five Earth-observation tasks in two flight environments. The results demonstrate reliable onboard inference under resource-constrained conditions.	arXiv:2512.01181
2025	Geospatial Foundation Models to Enable Progress on Sustainable Development Goals	This work introduces SustainFM, an SDG-grounded benchmark spanning 17 goals and 47 tasks to evaluate geospatial foundation models beyond raw accuracy. It emphasizes transferability, generalization, and energy efficiency, and motivates impact-driven and ethically informed deployment.	arXiv:2505.24528
2025	REOBench: Benchmarking Robustness of Earth Observation Foundation Models	Introduces REOBench, a robustness benchmark with six Earth-observation tasks and 12 corruption types spanning appearance and geometric shifts. Evaluations reveal substantial degradation in existing models and show comparatively stronger robustness from vision-language pretraining.	arXiv:2505.16793
2025	TerraFM: A Scalable Foundation Model for Unified Multisensor Earth Observation	Presents TerraFM, a global self-supervised multisensor model trained on Sentinel-1 and Sentinel-2 with large spatial tiles and land-cover-aware sampling. With modality-specific embeddings, cross-attention fusion, and tailored objectives, it reports strong gains on GEO-Bench and Copernicus-Bench.	arXiv:2506.06281
2025	Earth AI: Unlocking Geospatial Insights with Foundation Models and Cross-Modal Reasoning	Introduces Earth AI, a family of multimodal geospatial models integrating planet-scale imagery, population, and environmental data with a Gemini-based reasoning engine. A new benchmark based on real-world crisis scenarios shows cross-modal integration improving geospatial reasoning over single-modality baselines.	arXiv:2510.18318
2025	The Transparent Earth: A Multimodal Foundation Model for the Earth's Subsurface	Presents The Transparent Earth, a transformer foundation model for subsurface prediction across eight modalities using positional/modality encodings and text-derived embeddings. It supports inference with complete, partial, or no inputs and reports up to a threefold reduction in stress-angle prediction error versus prior methods.	arXiv:2509.02783
2025	Towards a Unified Copernicus Foundation Model for Earth Vision	Proposes a unified Copernicus foundation model trained on an 18.7M-image dataset spanning major Sentinel missions and both surface and atmospheric domains. The architecture combines dynamic hypernetworks with metadata encoding and demonstrates transfer across 15 Earth-vision tasks.	arXiv:2503.11849
2025	Foundation Models for Environmental Science: A Survey of Emerging Frontiers	Provides a survey of foundation models for environmental science, organizing applications such as forward prediction, data generation, data assimilation, downscaling, inverse modeling, ensembling, and decision-making. It also reviews data, architecture, training, and evaluation pipelines and identifies open research directions.	arXiv:2504.04280
2024	PANGAEA: A Global and Inclusive Benchmark for Geospatial Foundation Models	Presents PANGAEA, a standardized benchmark protocol for geospatial foundation models covering diverse datasets, tasks, sensors, resolutions, and temporal settings. Comparative results indicate current GFMs do not consistently outperform supervised baselines, especially under limited-label regimes.	arXiv:2412.04204
2024	Foundation Models for Remote Sensing and Earth Observation: A Survey	Surveys foundation models for remote sensing and Earth observation, including vision foundation models, vision-language models, LLM-related methods, and broader multimodal trends. It synthesizes public datasets and benchmarks, and discusses key challenges and future directions.	arXiv:2410.16602
2024	Exploring Foundation Models in Remote Sensing Image Change Detection: A Comprehensive Survey	Provides a systematic review of foundation-model approaches for remote-sensing image change detection, summarizing recent methods and framing key opportunities and challenges for this task.	arXiv:2410.07824