Impact of Systematic Relaxation Experiments on Subseasonal Forecast Skill in Machine Learning Weather Models

Forecasting extratropical weather on subseasonal timescales continues to be a challenge. One possible source of atmospheric predictability on these timescales are slowly evolving components of the climate system, most notably tropical modes of variability such as the Madden–Julian Oscillation (MJO) and tropical waves. These sources of predictability are not fully exploited because of systematic errors in numerical weather prediction (NWP) models. In particular, forecast errors that develop in the tropics at lead times of several days grow up-scale, propagate and degrade forecast skill in the extratropics on subseasonal timescales. Regions of forecast errors that exert the strongest influence on extratropical forecast skill remain poorly identified. Relaxation experiments using NWP models provide a means to isolate these regions, but such experiments are computationally demanding. In this study, we employ machine learning–based weather prediction models to perform relaxation experiments across multiple tropical regions. Probabilistic forecasts are generated using perturbed initial conditions from the Ensemble of Data Assimilations (EDA) of the European Centre for Medium-Range Weather Forecasts (ECMWF). Reforecasts covering a five-year period (2020--2024) are used to systematically assess the impact of relaxation strategies and the role of tropical variability, with a particular focus on the MJO, on extratropical subseasonal forecast skill. A key-finding is that predictions of the negative phase of the North Atlantic Oscillation improve when relaxation is applied following MJO phases 6 and 7 at initial time. Rossby wave source diagnostics are examined to investigate the dynamical processes leading to improvements in extratropical forecasts. The results demonstrate the value of relaxation experiment as a diagnostic tool when integrated with emerging machine learning–based prediction systems.