Deep DiveAI & Machine LearningMachine/Deep Learning

The Foundation Model That Learned Earth's Climate System

A widely discussed AI paper in Nature is not about language or imagesโ€”it's about Earth. This foundation model learns to predict weather, climate, and extreme events from a unified representation of the planet's physical systems.

By Sean K.S. Shin
This blog summarizes research trends based on published paper abstracts. Specific numbers or findings may contain inaccuracies. For scholarly rigor, always consult the original papers cited in each post.

If there is a single paper that defines the ambition of AI in 2025, it is not about chatbots or image generators. It is about the planet itself. Bodnar et al.'s foundation model for the Earth system, published in Nature, demonstrates that a single neural network can learn to represent and predict the coupled dynamics of atmosphere, ocean, and land surfaceโ€”the interacting physical systems that determine our climate, our weather, and ultimately our survival.

This is not merely an engineering achievement. It is a conceptual departure from centuries of physics-based climate modeling, and its implicationsโ€”for climate science, disaster preparedness, and the politics of environmental predictionโ€”are profound.

A Methodological Shift: From Equations to Data

Traditional numerical weather prediction (NWP) and climate modeling solve partial differential equations that describe fluid dynamics, thermodynamics, and radiative transfer on a discretized grid. The Integrated Forecasting System at ECMWFโ€”the gold standardโ€”runs these equations on supercomputers, consuming millions of CPU hours per forecast cycle. The models are magnificent in their physical rigor and frustrating in their computational cost.

AI weather models invert this paradigm. Instead of encoding physical equations, they learn input-output mappings from historical observation data. Given the state of the atmosphere at time t, they predict the state at time t+ฮ”tโ€”not by solving physics but by recognizing patterns in decades of observations. The result: forecasts that are dramatically faster than traditional NWP, at comparable or superior accuracy for medium-range forecasting.

Bodnar et al. push this paradigm to its logical conclusion: a foundation model that learns not just atmospheric dynamics but the coupled Earth systemโ€”atmosphere, ocean, land surface, and their interactions. Previous AI weather models treated the atmosphere in isolation. But weather and climate are fundamentally coupled systems: ocean currents drive atmospheric circulation; land surface moisture feeds back into precipitation patterns; sea ice extent modulates polar amplification. A model that ignores these couplings cannot capture the dynamics that matter most for long-range prediction and climate projection.

Architecture and Scale

The technical architecture reflects the ambition. The model processes multi-variate, multi-resolution data streams:

  • Atmospheric variables: temperature, humidity, wind, pressure at multiple vertical levels
  • Ocean surface: sea surface temperature, sea ice concentration, wave height
  • Land surface: soil moisture, snow cover, vegetation state
  • Satellite observations: directly assimilated radiance measurements
These heterogeneous data streams are unified through a transformer-based architecture that learns cross-modal attentionโ€”how atmospheric conditions attend to ocean states, how land surface conditions inform precipitation predictions. The foundation model is pre-trained on decades of reanalysis data (ERA5) and fine-tuned for specific prediction tasks.

The validation against operational forecasts is the paper's strongest claim: the foundation model matches or exceeds ECMWF's IFS on standard verification metrics for 1-10 day forecasts, while running in seconds rather than hours.

Extreme Events: Where AI Climate Models Struggle

Camps-Valls et al. provide the essential counterpoint. Their comprehensive review of AI for extreme weather reveals a systematic challenge: AI models trained on historical data underpredict extreme events.

The reason is statistical. Extreme eventsโ€”category 5 hurricanes, unprecedented heat waves, thousand-year floodsโ€”are, by definition, rare in the training data. Standard loss functions (mean squared error) penalize these rare events less than common conditions, causing models to regress toward climatological averages precisely when accuracy matters most.

The review identifies several emerging directions, including physics-informed architectures that embed conservation laws, statistical methods for modeling rare event tails, and causal discovery approaches that encode known physical mechanisms as inductive biases. None of these solutions is fully mature. The honest assessment: AI weather models are excellent for routine forecasting and dangerously overconfident for the extreme events that cause the most damage.

The Land Surface Gap

Mbarak et al.'s NoahMP-AI highlights a specific weakness in current Earth system AI: land surface prediction during extremes. Their work enhances the Noah Multi-Physics land surface model with deep learning, focusing on soil moisture predictionโ€”a variable critical for drought monitoring, flood forecasting, and agricultural planning.

The key finding: standard land surface models (both physics-based and AI-based) fail dramatically during drought and flood events, precisely because these extreme conditions lie outside the training distribution. Their hybrid approachโ€”using physics-based models as a backbone and deep learning to correct systematic biasesโ€”improves extreme-event prediction while maintaining physical consistency.

This hybrid philosophyโ€”physics for structure, AI for correctionโ€”may represent the pragmatic path forward for Earth system modeling: neither purely data-driven nor purely equation-based, but a marriage of both paradigms that inherits the strengths of each.

Claims and Evidence

<
ClaimEvidenceVerdict
AI foundation models match NWP for medium-range weather forecastingBodnar et al. demonstrate parity with ECMWF IFS on standard metricsโœ… Strongly supported
AI models capture coupled Earth system dynamicsFoundation model processes atmosphere-ocean-land jointlyโœ… Supported (for short-range)
AI weather models handle extreme events wellCamps-Valls et al. document systematic underestimationโŒ Refuted for current approaches
Physics-AI hybrids outperform pure AI for extremesNoahMP-AI shows improvement on drought/flood predictionโœ… Supported (early evidence)
AI climate models can replace physics-based projections for century-scale predictionNo validation beyond medium-range forecastingโš ๏ธ Unsubstantiated

Open Questions

  • Century-scale reliability: Can foundation models trained on 40 years of reanalysis data make credible projections 80 years into the futureโ€”under climate conditions that have no historical analogue? This is the central question for climate policy, and it remains unanswered.
  • Interpretability for scientists: Climate scientists need to understand why a model predicts what it predictsโ€”not just to build trust, but to advance physical understanding. Can AI Earth system models generate scientific insight, or only predictions?
  • Resolution limits: The foundation model operates at ~25km resolutionโ€”sufficient for synoptic-scale weather but inadequate for local impacts. Downscaling to km-scale for urban planning and agricultural applications remains an open challenge.
  • Ensemble prediction: Traditional NWP generates ensembles of forecasts to quantify uncertainty. How should foundation models generate calibrated ensemble predictions? The stochastic methods used in image diffusion models may provide a path, but the translation is non-trivial.
  • Governance of prediction: If AI weather models become dominant, who controls the infrastructure? Climate prediction is a public goodโ€”but foundation models are expensive to train and may concentrate prediction capability in the hands of a few well-resourced organizations.
  • What This Means for Your Research

    For climate scientists, the message is not that AI replaces physicsโ€”it is that AI augments physics in ways that enable new science. The speed advantage enables experimental designs that were previously computationally prohibitive: generating thousands of climate scenarios, systematically perturbing initial conditions, or running regional downscaling in real time.

    For AI researchers, Earth system modeling provides a uniquely rich evaluation domain. Unlike language or image generation, where quality is subjective, weather forecasts have ground truth. Every prediction can be verified against what actually happenedโ€”providing the kind of rigorous, objective evaluation signal that most AI benchmarks lack.

    For policymakers, the foundation model is both a promise and a warning. The promise: better, faster, cheaper environmental prediction to inform climate adaptation. The warning: these models have known blind spotsโ€”precisely in the extreme events that drive the most damage and demand the most urgent policy response.

    The Earth system foundation model is, in the most literal sense, a model of our world. Getting it right is not an academic exercise. It is a prerequisite for navigating the planetary crisis that defines this century.

    References (3)

    [1] Bodnar, C., Bruinsma, W., Lucic, A. et al. (2025). A foundation model for the Earth system. Nature.
    [2] Camps-Valls, G., Fernandez-Torres, M., Cohrs, K. et al. (2025). Artificial intelligence for modeling and understanding extreme weather and climate events. Nature Communications.
    [3] Mbarak, M., Singh, M., Sudharsan, N. et al. (2025). Towards NoahMP-AI: Enhancing Land Surface Model Prediction with Deep Learning. (Preprint; identifier pending).

    Explore this topic deeper

    Search 290M+ papers, detect research gaps, and find what hasn't been studied yet.

    Click to remove unwanted keywords

    Search 8 keywords โ†’