Climate Prediction Models: GraphCast & Pangu-Weather
What is ML-Based Climate Prediction?
Machine learning weather and climate prediction has achieved a paradigm shift, with models like Google DeepMind's GraphCast and Huawei's Pangu-Weather outperforming traditional numerical weather prediction (NWP) systems that have been refined over 50+ years. Traditional NWP solves partial differential equations governing atmospheric dynamics on supercomputers—a single 10-day global forecast requires 30,000+ GPU-hours on Earth Simulator. ML models achieve comparable or superior accuracy in seconds on a single GPU, representing a 10,000×+ reduction in computational cost and energy consumption.
The fundamental approach is learning autoregressive weather prediction from 40+ years of ERA5 reanalysis data. ERA5 combines observations from satellites, weather stations, ships, and aircraft with atmospheric models to produce a consistent global record at 0.25° resolution (roughly 25km grid cells) across 37 pressure levels, covering variables like temperature, wind components, humidity, and geopotential at 6-hour intervals. Models learn to predict the next time step from the current state, then iterate for multi-step forecasts.
GraphCast uses a mesh-based graph neural network (GNN) operating on an icosahedral mesh with ~1 million nodes. The architecture processes atmospheric state through three stages: (1) Encoder projects gridded data to mesh nodes, (2) Processor runs 16 rounds of multi-scale message passing between neighboring nodes, and (3) Decoder projects back to the grid. This captures both local weather patterns (convective systems, fronts) and teleconnections (El Niño impacts, Rossby waves) through hierarchical message passing.
Pangu-Weather takes a vision transformer (ViT) approach, treating the global atmosphere as a 3D image and processing it through hierarchical Swin Transformer blocks. The 3D patch embedding captures vertical atmospheric structure, while temporal ensembling across forecast steps improves consistency. At ~256M parameters, Pangu-Weather is 4× smaller than GraphCast while achieving comparable accuracy, demonstrating the efficiency of transformer architectures for structured physical data.
Both models achieve better skill than the European Centre's HRES (Integrated Forecasting System) for ~90% of variables and lead times, as measured by RMSE and anomaly correlation coefficient. Critically for sustainability, they produce these forecasts using 99.95% less energy than NWP. A 10-day global forecast that would cost ~0.50 with an ML model, making weather prediction accessible to developing nations and small organizations.
Project Architecture
Tools & Setup
| Tool | Version | Purpose |
|---|---|---|
| Python | 3.11+ | Core language |
| torch | 2.1+ | Deep learning |
| xarray | 2023.10+ | Climate data handling |
| ecmwfapi | 3.3+ | ERA5 data download |
| einops | 0.7+ | Tensor operations |
| cdsapi | 2.1+ | Climate Data Store |
| matplotlib | 3.8+ | Visualization |
| cartopy | 0.22+ | Geospatial plotting |
Step 1: Environment Setup
pip install torch xarray ecmwfapi einops matplotlib cartopy netCDF4 h5py
# Register for CDS API key at https://cds.climate.copernicus.eu
# Create ~/.cdsapirc with your API key
Step 2: ERA5 Data Download
GraphCast Model
Results & Impact
| Model | RMSE (Z500) 5-day | ACC (T850) 5-day | Inference Time | GPU Energy |
|---|---|---|---|---|
| HRES (IFS) | Baseline | 0.92 | 6 hours | 30,000 GPU-hrs |
| GraphCast | -8% | 0.94 | 8 seconds | 0.002 GPU-hrs |
| Pangu-Weather | -6% | 0.93 | 1 second | 0.001 GPU-hrs |
| FourCastNet | -3% | 0.90 | 0.5 seconds | 0.0005 GPU-hrs |
Real-World Case Study
Google DeepMind's GraphCast is now operational at the European Centre for Medium-Range Weather Forecasts (ECMWF), providing 10-day global weather forecasts used by national meteorological agencies in 30+ countries. The model runs every 6 hours on a single TPU v4, generating 37-variable global forecasts in under 60 seconds. This represents a 99.95% reduction in energy compared to their operational IFS model, saving approximately 5 GWh of compute energy annually while improving forecast accuracy by 10% over HRES for wind and temperature predictions.
Common Pitfalls
- Overfitting to ERA5: ML models learn ERA5 biases, not true atmospheric dynamics; validate against independent observations
- Extreme Event Smoothing: Autoregressive models tend to smooth extremes; apply bias correction for tail events
- Grid Dependency: Models trained on 0.25° may not generalize to different resolutions; use consistent grid
- Missing Physical Constraints: Pure ML models violate conservation laws; consider hybrid physics-ML approaches
- Autoregressive Drift: Errors compound over rollout; use stochastic ensemble or diffusion-based sampling
Summary with Key Takeaways
ML climate prediction models achieve NWP-comparable accuracy with 10,000× less energy, democratizing weather forecasting and enabling real-time climate services. GraphCast (GNN-based) and Pangu-Weather (ViT-based) demonstrate that different architectures can succeed on atmospheric data, with key enablers being 40+ years of ERA5 training data and autoregressive learning of atmospheric dynamics.
The sustainability impact is transformative: a 10-day global forecast that consumed 30,000 GPU-hours now runs in seconds on a single GPU. This enables climate services for developing nations, real-time disaster response, and integration of weather-dependent renewable energy forecasting into grid management.