Divergent European hail projections from machine learning and physically based models under global warming

Reliable fine-scale hail projections are urgently needed for robust risk assessments, yet future trends remain uncertain and contradictory across methods. Using 11-year pan-European convection-permitting climate simulations for the present and a + 3°C pseudo-global-warming climate, we compare an online hail-growth diagnostic (HAILCAST) with an offline machine learning model (XGBoost) trained on ERA5 hail environments. We show that, under current conditions, both approaches produce comparable hail frequencies across most of Europe. However, XGBoost predicts widespread hail suppression in a warmer climate, mainly driven by increasing freezing level heights, from conditions beyond the training distribution. HAILCAST instead simulates how enhanced storm updrafts can sustain hail growth despite a warmer atmosphere, projecting regional increases over central-eastern Europe and larger hailstones. Our findings show that data-driven approaches alone should be used with caution under global warming, underscoring the need for physically based hail representations in convection-permitting climate models to robustly assess future convective hazards.