Relaxation experiments in ML-based weather prediction models to study subseasonal predictability

Abstract. This study explores the use of relaxation experiments in machine learning-based weather prediction (MLWP) models to identify sources of subseasonal predictability in comparison to a traditional numerical weather prediction (NWP) system. Relaxation involves nudging specific regions of a model toward reanalysis data to isolate their influence on forecast skill. We apply this technique to two MLWP models, Pangu-Weather (fully data-driven) and NeuralGCM (hybrid) and compare the experiments to the Unified Forecast System (UFS). The focus is on week 3–4 forecast of two major precipitation events in western North America in winter 2022/2023, both linked to Madden-Julian Oscillation (MJO) activity. For the two cases, MLWP models exhibit higher forecast skill than the UFS at subseasonal lead times. Though tropical relaxation improves the skill in all forecast systems, gains are greater for UFS, reflecting the MLWP models’ stronger baseline performance. A Rossby wave source (RWS) analysis shows that tropical relaxation consistently improves the large-scale dynamic processes associated with the tropical-extratropical teleconnections leading to both events. These results highlight the potential of relaxation experiments as a low-cost, effective diagnostic for understanding and improving subseasonal forecasts, especially in emerging MLWP systems.