Bridging Artificial Intelligence and Data Assimilation: The Data-driven Ensemble Forecasting System ClimaX-LETKF

While machine learning-based weather prediction (MLWP) has achieved significant advancements, research on assimilating real observations or ensemble forecasts within MLWP models remains limited. We introduce ClimaX-LETKF, the first purely data-driven ML-based ensemble weather forecasting system. It operates stably over multiple years, independently of numerical weather prediction (NWP) models, by assimilating the NCEP ADP Global Upper Air and Surface Weather Observations. The system demonstrates greater stability and accuracy with relaxation to prior perturbation (RTPP) than with relaxation to prior spread (RTPS), while NWP models tend to be more stable with RTPS. RTPP replaces an analysis perturbation with a weighted blend of analysis and background perturbations, whereas RTPS simply rescales the analysis perturbation. Our experiments reveal that MLWP models are less capable of restoring the atmospheric field to its attractor than NWP models. This work provides valuable insights for enhancing MLWP ensemble forecasting systems and represents a substantial step toward their practical applications.

💡 Research Summary

This paper introduces ClimaX‑LETKF, the first purely data‑driven ensemble weather forecasting system that assimilates real observations without relying on any conventional numerical weather prediction (NWP) model for generating initial conditions. The system couples the low‑resolution (5.625°) ClimaX deep‑learning model—trained from scratch on the WeatherBench subset of ERA5—with a Local Ensemble Transform Kalman Filter (LETKF) that uses 20 ensemble members. Real‑world observations from the NCEP ADP Global Upper‑Air and Surface Weather Observations are ingested every six hours, pre‑processed (unit conversion, thinning, error assignment) and quality‑controlled (gross‑error check).

Two covariance inflation strategies are examined: Relaxation to Prior Spread (RTPS) and Relaxation to Prior Perturbation (RTPP). Both methods employ a tunable relaxation parameter α. In RTPS, analysis perturbations are rescaled to match the prior (background) ensemble spread; in RTPP, analysis perturbations are replaced by a weighted blend of analysis and background perturbations. Experiments covering a two‑year DA cycle (January 2016–December 2017) reveal markedly different behavior. With RTPS, the system becomes unstable for several α values (e.g., α = 1.3) showing sharp RMSE spikes in early 2016, suggesting that the instability stems from physical imbalance rather than under‑dispersion. In contrast, RTPP with α ≈ 0.90 yields the lowest and most stable RMSE, and the ensemble spread stays close to the RMSE throughout the period. This indicates that the data‑driven model is less chaotic than traditional NWP models and requires larger inflation factors to compensate for model imperfections.

Spatial diagnostics show that background RMSE and spread patterns are similar across temperature (500 hPa), zonal wind (700 hPa), meridional wind (850 hPa), and surface pressure. Excessive spread appears in low‑ to mid‑latitude regions, Greenland, and Antarctica, while high‑latitude oceanic areas exhibit under‑dispersion. A negative Pearson correlation between error magnitude and observation density (−0.31 to −0.45) confirms that denser observations improve analysis quality, especially in the Northern Hemisphere low latitudes. Large surface‑pressure errors over Russia and northern Canada are attributed to mismatches between the model’s coarse topography and real terrain, despite abundant observations.

Five‑day ensemble forecasts (6‑hour steps) using the optimal inflation settings demonstrate that RTPP maintains a spread that tracks the RMSE more closely than RTPS, which initially reduces spread at 6 h before it grows again. Nevertheless, the spread growth is slower than the RMSE increase, reflecting the reduced chaotic nature of the machine‑learning model compared with physics‑based NWP systems.

The study concludes that a purely data‑driven ensemble system can stably assimilate real observations over multi‑year periods, but its performance and stability are highly sensitive to the choice and tuning of covariance inflation. RTPP proves more suitable for the ClimaX model, whereas traditional NWP systems usually favor RTPS. The findings highlight the need for stronger physical constraints, higher‑resolution observations, and larger ensembles to improve uncertainty representation in future ML‑based forecasting frameworks.