AI diffusion model beats major weather ensembles on North Pacific water vapor forecasts

The diffusion ensemble outperformed the operational GEFS and ECMWF ensembles on continuous ranked probability score, spread-to-error ratio, reliability diagrams and Brier skill scores. Once trained, it produces forecasts quickly and at low computational cost.

Traditional ensembles perturb initial conditions and model physics to represent uncertainty. This approach replaces that step with a trained generative model, which matters for operations that depend on risk assessment. The abstract reports results for medium-range IVT forecasts only.

The study replaces the usual way of building ensembles, which slightly perturbs initial conditions and the numerical solvers and physics of a model to capture uncertainty. Instead, a diffusion model is trained to generate many plausible forecasts from a single deterministic West-WRF run.

Evaluation used binned ranked histograms, continuous ranked probability score, binned spread-to-root-mean-square-error ratio, reliability diagrams and Brier skill scores. On these measures the diffusion ensemble outperformed the operational GEFS and ECMWF ensembles.

The authors also report that the ensemble represented real outcomes of extreme destructive events with realistic, high-quality images, suggesting use in risk assessment for operations that rely on ensemble forecasts.