AI for Weather Forecasting

Weather forecasting has been one of the great success stories of computational science: modern 5-day forecasts are more accurate than 1-day forecasts were 40 years ago, driven by increasingly accurate numerical weather prediction (NWP) models that simulate atmospheric physics on supercomputers. In 2023, a wave of AI-based forecasting models dramatically challenged NWP supremacy — matching or exceeding the accuracy of ECMWF’s Integrated Forecasting System (IFS), the world’s best operational NWP model, at a fraction of the computational cost.

Background: Numerical Weather Prediction

Traditional NWP works by numerically integrating the equations of atmospheric dynamics (Navier-Stokes, thermodynamics, continuity) on a 3D global grid, updated every 6–12 hours with assimilated observations from satellites, radiosondes, and surface stations. ECMWF’s IFS runs at ~9 km horizontal resolution with 137 vertical levels and requires millions of CPU-hours per forecast. While physically principled and interpretable, NWP is:

• Computationally expensive: a 10-day ensemble forecast costs ~10,000 CPU-core-hours.
• Slow to produce: the full forecast pipeline takes 1–2 hours of wall-clock time.
• Constrained by the parameterization problem: subgrid-scale processes (convection, cloud microphysics) must be approximated.

AI models learn the mapping from an initial atmospheric state directly to future states, bypassing explicit physics simulation.

The Reanalysis Training Paradigm

All major AI weather models are trained on ERA5 — ECMWF’s fifth-generation atmospheric reanalysis, covering 1979–present at 0.25° resolution (~28 km), 37 pressure levels, and hourly frequency. ERA5 contains ~6.5 petabytes of gridded data representing temperature, wind, geopotential, humidity, and more at each grid point.

The standard supervised learning setup:

• Input: Global atmospheric state at time $t$ (and usually $t - 6h$ for velocity tendencies).
• Target: Atmospheric state at time $t + \Delta t$ (one step ahead).
• Autoregressive rollout: Iteratively apply the model to produce forecasts at $t, t+\Delta t, t+2\Delta t, \ldots$

The models are evaluated against ECMWF IFS on the WeatherBench 2 benchmark.

GraphCast (Google DeepMind, 2023)

GraphCast (Lam et al., 2023) uses a Graph Neural Network over a multi-scale icosahedral mesh to model atmospheric dynamics.

GraphCast Architecture

• Grid-to-mesh encoder: Projects ERA5 grid points onto graph nodes in an icosahedral mesh (O(40,000 nodes) at ~1° resolution).
• Processor: 16 layers of message passing on the mesh, with edges representing both local (short-range) and long-range connections across multiple mesh resolutions.
• Mesh-to-grid decoder: Projects mesh node features back to the 0.25° ERA5 grid.

The multi-resolution mesh (6 levels of icosahedral refinement) allows GraphCast to efficiently propagate information across both local and planetary scales — crucial for weather phenomena that span continents.

GraphCast Performance

GraphCast achieves better skill than ECMWF IFS on 90% of 1,380 verification targets (all pressure levels, all variables, all lead times from 6h to 10 days) on the HRES analysis. It is particularly strong at:

• Medium-range forecasting (3–7 days): 500 hPa geopotential height, 850 hPa temperature.
• Tropical cyclone tracking: 15–20% improvement in track error vs. IFS.
• Extreme event 2m temperature forecasts.

Pangu-Weather (Huawei, 2023)

Pangu-Weather (Bi et al., 2023) uses a 3D Earth attention transformer trained to minimize a weighted $\ell_1$ + $\ell_2$ loss on ERA5.

Pangu-Weather Architecture

A hierarchical transformer with:

• 3D window attention: Attention computed within overlapping 3D windows in the (latitude, longitude, pressure level) space, capturing local vertical and horizontal correlations.
• Shifted windows: Analogous to Swin Transformer — alternating attention windows across layers provide cross-window communication.
• Multi-scale design: Four separate models for forecast lead times of 1h, 3h, 6h, and 24h; at inference, models are combined to produce arbitrary-horizon forecasts with minimal error accumulation.

The multi-resolution model strategy (separate networks per horizon) reduces autoregressive error accumulation: instead of rolling out 40 consecutive 6-hour steps to reach 10 days, Pangu-Weather can use fewer, larger steps.

Pangu-Weather Performance

Pangu-Weather surpasses ECMWF IFS HRES on nearly all upper-air variables at forecast horizons of 1–7 days, with particular strength on upper-level wind forecasts.

FourCastNet (NVIDIA, 2022)

FourCastNet (Pathak et al., 2022) was the first AI model to demonstrate competitive NWP skill at global resolution, predating GraphCast and Pangu-Weather. It uses an Adaptive Fourier Neural Operator (AFNO) backbone:

• The global atmospheric state is processed as a 2D image at each pressure level.
• AFNO applies token mixing in Fourier space — efficient global mixing at $O(n \log n)$ cost.
• FourCastNet generates a 10-day forecast in under 2 seconds on a single A100 GPU.

FourCastNet enabled large ensemble forecasting at previously unaffordable scale: generating 1,000-member ensembles to characterize forecast uncertainty, vs. the 51-member operational ECMWF ensemble.

NeuralGCM (Google, 2024)

NeuralGCM (Kochkov et al., 2024) represents a hybrid approach: combining a differentiable atmospheric dynamical core (physics equations for large-scale dynamics) with neural network parameterizations for subgrid processes.

Unlike pure ML models, NeuralGCM:

• Enforces physical conservation laws (mass, energy, momentum) exactly through the dynamical core.
• Uses learned neural networks only for the parameterization residuals (convection, cloud processes) that NWP traditionally approximates with handcrafted schemes.
• Is end-to-end differentiable — the full hybrid model can be trained with gradient descent.

NeuralGCM outperforms both pure ML models and IFS at multi-week horizons (beyond 10 days) where ML models’ error accumulation begins to dominate. It also produces better calibrated uncertainty estimates.

Aurora (Microsoft, 2024)

Aurora is a large foundation model (1.3B parameters) pretrained on a diverse corpus of atmospheric data: ERA5, CMIP6 climate model outputs, MERRA-2 reanalysis, HRES analysis, and regional high-resolution datasets. Key features:

• 3D Swin Transformer backbone operating on pressure-level and surface variables jointly.
• Variable-resolution pretraining: Trained on data at multiple resolutions (0.1° to 1°), enabling zero-shot adaptation to new grids.
• Multistep fine-tuning: Fine-tuned with autoregressive rollout loss over multiple steps to improve long-range stability.

Aurora achieves state-of-the-art performance on AIFS (ECMWF’s operational AI model) benchmarks at 0.1° resolution, with strong performance on air quality forecasting and ocean wave height in addition to standard atmospheric variables.

Skill Metrics

AI weather models are evaluated using standard meteorological skill metrics:

The 500 hPa geopotential height Z500 is the canonical benchmark variable: it governs large-scale weather patterns and is reliably observed and analyzed. All major AI models achieve ACC > 0.6 out to ~10 days on Z500 — matching or exceeding IFS.

Advantages and Limitations of AI Weather Models

Advantages

• Speed: Generate a 10-day global forecast in seconds on a single GPU vs. hours on a supercomputer.
• Cost: $100–$1,000 per forecast vs. $100,000+ for operational NWP.
• Ensemble generation: Affordable 1,000+ member ensembles for improved uncertainty quantification.
• Extreme event skill: Some AI models show improved skill on rare extremes vs. IFS.

Limitations

• Blurry high-resolution predictions: Autoregressive ML models tend to produce spatially smooth (blurry) precipitation forecasts at fine scales — failing to capture extreme local intensity.
• Physical consistency: AI forecasts may violate conservation laws (mass, energy); wind and temperature fields may be locally inconsistent.
• Training data ceiling: All models are trained on ERA5, which itself has biases and limited resolution. Models cannot generalize beyond ERA5’s quality.
• Rare events: Extremely rare events (once-per-century storms) are poorly represented in 40 years of training data.
• Climate change extrapolation: ML models trained on historical data may not extrapolate to a warming climate that diverges from historical distributions.

The Path to Operational Deployment

ECMWF launched its own AI model AIFS (Artificial Intelligence Forecasting System) in 2024, integrating an AI model into operational forecasting alongside IFS. The hybrid approach:

1. AI model provides fast, accurate deterministic and ensemble forecasts.
2. IFS provides physically consistent, high-resolution forecasts for safety-critical decisions.
3. AI models are used for post-processing — correcting systematic biases in NWP output (Model Output Statistics, MOS).

National weather agencies (NOAA, ECMWF, MetOffice) are actively testing AI models for operational use, with full AI-primary forecasting expected to begin by 2027.

Summary

AI weather foundation models — GraphCast, Pangu-Weather, FourCastNet, NeuralGCM, and Aurora — have demonstrated that machine learning trained on ERA5 reanalysis can match or exceed the world’s best NWP models at medium-range forecasting, at 10,000× lower compute cost. GNN-based architectures (GraphCast) and hierarchical transformers (Pangu-Weather, Aurora) capture planetary-scale dynamics; hybrid physics-ML models (NeuralGCM) combine physical conservation with learned parameterizations. Key remaining challenges are high-resolution precipitation, physical consistency, and extrapolation to climate change scenarios.