PIKA : A Physics-Informed Koopman Autoencoder for Glacier Mass Balance Forecasting

Anthropogenic warming is driving sustained glacier mass loss worldwide, contributing roughly one-fifth of observed sea-level rise this century and eroding the seasonal meltwater buffer that mountain agriculture, hydropower, and municipal supply depend on.Strategic resource planning is undermined by the dual constraints of limited observational data and short forecast lead times We present PIKA, a compact model that predicts annual glacier mass balance from tabular climate and glacier inventory records, without satellite imagery.Across a global benchmark of 121 glaciers, our architecture matches a standard deep learning model $9\times$ its size in distribution, supplies conformally calibrated prediction intervals, and recovers a single interpretable memory timescale of about three and a half years, consistent with how long compacted snow retains a climate signal. Seed-paired analyses attribute the achieved accuracy to the holistic compact framework rather than individual components, establishing in situ data sparsity over model capacity as the critical limiter for multi-year forecasting horizons. These findings indicate that lightweight, calibrated architectures constrained by standard observational datasets represent a viable approach for supporting long-term water and energy planning.