Model Deep Dives | DeepMind Weather AI

DETERMINISTIC

GraphCast

Learning skillful medium-range global weather forecasting • Science 2023

📊 Specifications

Parameters36.7M

Resolution0.25° (~25km)

Timestep6 hours

Forecast Horizon10 days (40 steps)

Variables227 (37 levels)

Mesh Nodes40,962 (M6)

Inference Time<60 sec (TPU)

🏗️ Architecture

GraphCast uses an encode-process-decode GNN architecture operating on a multi-scale icosahedral mesh.

// High-level architecture
Encoder:  lat-lon grid (721×1440) → mesh (40,962 nodes)
Processor: 16 layers of message passing on multi-mesh
Decoder:  mesh → lat-lon grid (prediction Δ)

// Message passing (per layer)
edges' = MLP_e(concat(edges, nodes[sender], nodes[receiver]))
nodes' = MLP_v(concat(nodes, aggregate(edges')))
                        

Key Innovation: Multi-mesh retains edges from all refinement levels (M0-M6), enabling both local detail and global information flow in a single message-passing step.

📈 Performance vs HRES

Outperforms on 90% of 1380 verification targets
99.7% better in troposphere (most important region)
Superior tropical cyclone track prediction
Better atmospheric river and extreme temperature forecasts
Orders of magnitude faster (seconds vs hours)

📚 Resources

📄 arXiv Paper 🔬 Science Publication 💻 GitHub Repository

PROBABILISTIC

GenCast

Diffusion-based ensemble forecasting for medium-range weather • Nature 2024

📊 Specifications

Parameters~60M

Resolution0.25°

Timestep12 hours

Forecast Horizon15 days

Ensemble Members50+

Pressure Levels13

Inference Time8 min/member

🌀 Diffusion Process

GenCast is a diffusion model adapted to spherical Earth geometry. It learns to generate probabilistic weather trajectories by iteratively denoising.

// Forward process (training)
X_t = √ᾱ_t · X_0 + √(1-ᾱ_t) · ε    // Add noise

// Reverse process (inference)
for t = T to 1:
    ε_pred = denoiser(X_t, t, X^{t-1}, X^t)  // Predict noise
    X_{t-1} = denoise_step(X_t, ε_pred)       // Remove noise
                        

Key Innovation: Uses sparse transformer attention (neighbors only) in the processor, enabling efficient computation while capturing local atmospheric dynamics.

📈 Performance vs ENS

Outperforms on 97.4% of 1320 targets
99.8% better at lead times >36 hours
Superior tropical cyclone track prediction
Better extreme heat/cold/wind forecasting
More accurate wind power production forecasts

📚 Resources

📄 arXiv Paper 🔬 Nature Publication 💻 GitHub Repository

FUNCTIONAL GENERATIVE

FGN (WeatherNext 2)

Skillful joint probabilistic weather forecasting from marginals • 2025

📊 Specifications

Parameters180M (per seed)

Model Seeds4 (deep ensemble)

Latent Dimension768

Transformer Layers24

Timestep6 hours

Noise Dimension32

Speed vs GenCast8× faster

⚡ Functional Perturbation

FGN generates ensembles via learned functional perturbations — injecting low-dimensional noise directly into the architecture.

// Key innovation: 32D noise → global weather field
z ~ N(0, I)^32  // Sample 32-dimensional noise

for each layer:
    x = ConditionalLayerNorm(x, z)
    // γ(z)·(x-μ)/σ + β(z)
    // Same z applied globally → spatial coherence

// Training: minimize CRPS on marginals
// Result: learns realistic joint distributions!
                        

Key Insight: Trained only on per-location CRPS (marginals), but the shared global noise forces the model to learn realistic spatial correlations (joints).

📈 Performance vs GenCast

Better on 99.9% of variables and lead times
~6.5% average CRPS improvement
~24 hours earlier cyclone track skill
8× faster inference (single pass vs iterative)
Powers Google Search, Maps, Pixel Weather

📚 Resources

📄 arXiv Paper 📝 Google Blog Post 🔧 WeatherNext API

SHARED INFRASTRUCTURE

Icosahedral Mesh

The spherical representation powering all three models

🌐 Why Icosahedral?

Traditional lat-lon grids have fundamental problems for global weather modeling:

Polar singularity: Grid cells shrink to points at 90°N/S
Aspect ratio: Cells near poles extremely elongated
Uneven area: Polar cells much smaller than equatorial
CFL constraint: Tiny polar cells limit timestep globally

Icosahedral meshes solve all of these with near-uniform node spacing.

📐 Refinement Levels

Level	Nodes	Resolution
M0	12	~7,000 km
M1	42	~3,500 km
M2	162	~1,750 km
M3	642	~875 km
M4	2,562	~440 km
M5	10,242	~220 km
M6	40,962	~110 km