Improving Subseasonal Weather Forecast Using Tropical Weighting: A Fine-Tuned 2D Transformer

Subseasonal forecasts, targeting lead times from about 2 weeks to 2 months, remain challenging. This time range between medium-range weather and seasonal climate predictions is often described as a “predictability desert”, where both numerical weather prediction (NWP) models and machine learning (ML)-based systems tend to lose skill or have not been rigorously evaluated. In this work, we fine-tune a 2D Transformer-based model derived from the Pangu-Weather architecture for 30-day subseasonal forecasts. The focus is on improving week-3 and week-4 lead times by assigning extra weights to the tropics, which host slowly varying modes of variability that influence global weather. During model training, we apply region-based weighting using a smooth Gaussian function centered at the equator. This function assigns higher weights to tropical latitudes, with the width of the weighting controlled by a tunable standard deviation parameter. The model is trained on a multi-year subset of 6-hourly ERA5 reanalysis data and uses five upper-air variables (geopotential, temperature, zonal and meridional wind components, specific humidity) at 13 pressure levels, along with four surface variables (mean sea-level pressure, 2-meter temperature, 10-meter winds), totaling 69 input channels. For inference, we generate both deterministic and ensemble forecasts. The deterministic forecasts are initialized using ERA5 reanalysis fields, while the ensemble forecasts use 10 perturbed members from ECMWF’s Ensemble Data Assimilation (EDA), enabling probabilistic forecast evaluation. Forecast evaluation is conducted using both deterministic and probabilistic metrics. Compared to the 2D Transformer baseline, the fine-tuned model shows approximately 70% bias reduction and up to 50% RMSE improvement for temperature (T850 and T2m), particularly at week-3 and week-4 lead times. CRPS scores also generally improve, indicating better ensemble skill and reliability.