Weather & Earth Science: The Revolution of Forecast

Bottom line: Deep-learning medium-range forecasting is not "a faster numerical forecast" — it is an independent technical route that learns atmospheric evolution directly from massive reanalysis data, compressing a 10-day forecast from "hours on a supercomputer" to "seconds on a single GPU". GraphCast, Pangu, and FourCastNet are three representative routes; what actually matters in production is data conventions, rolling inference, and interpretability validation.

1. The Medium-Range Problem: Why Rework It with AI

Numerical weather prediction (NWP) is among the largest scientific computing workloads on Earth:

• Cost: ECMWF's IFS runs on a ~100-PFlops-class supercomputer; one 10-day global forecast takes hours;
• Resolution: the 0.25° global grid is ~1440×721 cells × multiple variables and levels; a single forecast produces tens of GB;
• Bottleneck: approximation error of parameterization schemes (convection, radiation, cloud microphysics) + exponential growth of compute cost.

The AI route does not aim to replace NWP but to provide millisecond inference at comparable accuracy, plus far cheaper ensembles (50 perturbed members for the cost of one run, quantifying uncertainty).

2. GraphCast: Equivariant Message Passing on Graphs

DeepMind's GraphCast core idea: discretize the atmosphere onto a multi-resolution graph and run equivariant message passing with graph networks.

• Grid: hexagonal multi-grid (0.25° and 1° levels), progressively upsampling/aggregating from fine to coarse;
• Training: on ERA5 reanalysis — input 10 days of history, predict 6-hour-later states;
• Inference: autoregressive rollout — predict 6 hours ahead each step, feed predictions back, roll 40 steps for a 10-day forecast;
• Results: beats ECMWF's deterministic high-resolution forecast (IFS HRES) on 90%+ of variables, at seconds per inference on a single TPU/GPU.

Engineering key: error accumulation in rollout is the biggest risk — a small per-step error does not mean 40 steps remain usable; drift and spectral-energy degradation must be monitored per step.

3. Pangu: 3D Earth-Specific Transformer

Huawei Cloud Pangu-Weather's differentiating design:

• 3D Earth-Specific Transformer: treats the atmosphere as 3D volume data (lon × lat × height × variables), attending within spherical local windows — avoiding planar-Transformer distortion at the poles;
• Hierarchical temporal aggregation: multiple temporal-aggregation modules predict 1/3/6/24-hour targets in parallel and fuse them — lower rollout error than pure autoregression;
• 4D data: pressure-level multi-variable + surface variables as input; hourly (or 3-hourly) output for 10 days;
• Advantage: leading medium-range accuracy with far lower inference latency than NWP.

Engineering point: parallel multi-horizon prediction costs more memory/compute than single-horizon autoregression but buys lower rollout error — compute for accuracy, suited to accuracy-first production.

4. FourCastNet: Fourier Neural Operators and Generative Forecasting

NVIDIA's FourCastNet takes the neural-operator route:

• AFNO (adaptive Fourier neural operator): global convolution in spectral space, naturally capturing large-scale teleconnections and multi-scale structure;
• Training: 0.25° global ERA5, 20 hours of history in, 6-hour forecast out;
• Generative extension: diffusion/generative training produces ensemble forecasts — multiple perturbed members per run, quantifying probabilities directly;
• Positioning: fastest; ideal for ensembles and super-resolution; absolute accuracy slightly behind GraphCast/Pangu.

5. Comparing the Three Routes

Dimension	GraphCast	Pangu	FourCastNet
Core architecture	Multi-resolution graph network	3D spherical Transformer	Fourier neural operator
Inference mode	Autoregressive rollout	Multi-horizon fusion	Single-step + generative ensembles
Accuracy	Leading overall	Leading medium-range	Moderate
Speed	Seconds	Seconds	Fastest; ensemble-ready
Best fit	Deterministic baseline	Accuracy-first production	Ensembles / super-resolution

6. Engineering Checklist

• Data conventions: full ERA5 reanalysis is PB-scale — slice by variable/time into TFRecord/NetCDF shards; normalize with climatological statistics, never batch statistics;
• Rollout engineering: autoregression is a "loop calling the model" — manage state (write predictions back into input tensors) and monitor errors per variable (RMSE, energy spectrum);
• Compute profile: weeks on hundreds of GPUs for training; seconds on one GPU for inference — the same resource shape as SFM training/inference;
• Validate by climate zone: global mean metrics hide polar/tropical bias; evaluate in latitude bands;
• Mix with NWP: use AI forecasts as fast pre-screens or ensemble members; extreme events (typhoon tracks) must be re-checked with traditional models.

Success criterion: long-term verifiability. An AI weather model must be continuously back-tested against real observations (not just reanalysis) after release, with per-day rollout error monitored — one quarter of independent validation beats any release-day accuracy table.