WeatherMesh-3: Fast and accurate operational global weather forecasting

TL;DR

WM-3 tackles the challenge of delivering fast, accurate global weather forecasts with limited hardware by introducing a latent-rollout transformer framework that preserves latent-space continuity across forecast steps. The encoder–processor–decoder architecture, powered by NATTEN neighborhood attention and rotary embeddings, enables arbitrary lead-time predictions through repeated latent-space steps, while pretraining on ERA-5 and operational fine-tuning with IFS/GFS analyses support real-time use. Empirically, WM-3 achieves state-of-the-art accuracy relative to operational models (e.g., up to $37.7\%$ RMSE improvement for $2$‑meter temperature at 1 day) and reduces forecast blur, all while delivering a $14$‑day forecast in $12$ seconds on a single RTX $4090$ and running on consumer-grade hardware. The work also emphasizes accessibility and extensibility, with modular encoders for additional data sources and open-source tooling to facilitate broader deployment and future work in data assimilation and ensemble methods.

Abstract

We present WeatherMesh-3 (WM-3), an operational transformer-based global weather forecasting system that improves the state of the art in both accuracy and computational efficiency. We introduce the following advances: 1) a latent rollout that enables arbitrary-length predictions in latent space without intermediate encoding or decoding; and 2) a modular architecture that flexibly utilizes mixed-horizon processors and encodes multiple real-time analyses to create blended initial conditions. WM-3 generates 14-day global forecasts at 0.25-degree resolution in 12 seconds on a single RTX 4090. This represents a >100,000-fold speedup over traditional NWP approaches while achieving superior accuracy with up to 37.7% improvement in RMSE over operational models, requiring only a single consumer-grade GPU for deployment. We aim for WM-3 to democratize weather forecasting by providing an accessible, lightweight model for operational use while pushing the performance boundaries of machine learning-based weather prediction.

Figures (9)

• Figure 1: (a) A schematic of the encoder-processor-decoder architecture. (b) Illustration of the difference in attention window location between SWIN and NATTEN.
• Figure 2: (a) Scorecard comparing WM-3 vs. IFS HRES across all pressure levels and surface variables. (b) Scorecard comparing WM-3 vs. AIFS.
• Figure 3: (a) Forecast maps showing 10-meter windspeed and mean sea level pressure at 72 hours lead time from 00z on July, 1, 2024. MLWP models offer notably less fine-grained spatial details, but resolve the intensity of Hurricane Beryl more clearly than the IFS ensemble mean. (b) Blur score vs RMSE for select surface variables. The points on the IFS Ensemble line represent ensemble means for member subsets beginning at 1 member and ending at 51 members.
• Figure 4: Schematic for operational encoders.
• Figure 5: RMSE-blur score for surface variables.