Calibrating AI Subseasonal to Seasonal(S2S) Forecasts with Lightweight Regional Adapters

Global AI weather models provide skillful broad-domain forecasts, but global skill does not guarantee reliable regional uncertainty estimates. Using weekly rainfall over India as a single-variable case study, we find that raw FuXi-S2S ensemble members are strongly underdispersed. We introduce a lightweight probabilistic adapter that post-processes the fixed ensemble without retraining its global generator. The adapter combines a shared member encoder, permutation-invariant pooling, and a regional convolutional network to estimate spatially varying, lead-dependent location and spread corrections while retaining member-index correspondence. In the primary retrospective evaluation, it reduces the continuous ranked probability score (CRPS) by 16.21% relative to raw FuXi and by a further 4.26% relative to structured moment calibration. Sensitivity analyses and an independent reconstruction of the scoring pipeline support this result, although interval coverage remains subnominal and bias improvement is unresolved. The evidence is limited to FuXi, rainfall, and India. Additional physically relevant predictors may provide richer regional context, while transfer to other regions must be evaluated rather than assumed.