Deep-learning forecast output from the DLSM experiments

These forecast experiments are built on the Deep Learning Earth System Model (DLESyM, Cresswell-Clay et al. 2025) and its stratospheric extension (DLSM). Both are trained on 33 years of ERA5 reanalysis (1983–2015) and validated on 2016–2017. DLSM operates on an HPX64 grid (~100 km) with a 48‑h time step, outputting geopotential height and temperature at 10 and 50 hPa, while the atmospheric component supplies coupling variables at 100, 500, and 850 hPa. To isolate stratospheric influence on surface wind predictability, we run twin atmospheric models: one with no input above 100 hPa (control) and one additionally ingesting 50‑hPa fields from DLSM. Both are coupled identically to the ocean module. Forecast outputs provided here are 10‑m total wind speed (ws10) and 500‑hPa geopotential height (z500).