Convection-permitting scales break climate invariance in machine-learned precipitation downscaling

Xingjian Yan, Anamitra Saha, Erin Dougherty, Sai Ravela

Machine-learning precipitation downscaling trains a coarse-to-fine mapping operator applicable to future climate projections, avoiding the cost of km-scale numerical simulations. This implicitly assumes the climate-change signal is contained in the coarse input and the learned operator remains nearly invariant. Testing this assumption over Puerto Rico with current (2001-2021) and near-future (2041-2061) extreme-rainfall resimulations, we find current- and future-climate operators nearly indistinguishable at 48-25 and 25-12 km but diverge at 12-3 km, where the current-climate operator underestimates future fine-scale variance. Kinetic and available potential energy budgets of idealized MPAS squall-line simulations explain this scale dependence: diabatic heating and buoyancy conversion increase disproportionately below 25 km and reorganize shell-to-shell kinetic-energy transfers toward finer scales. Climate non-stationarity is strongest at scales coarse climate models do not explicitly resolve. Km-scale downscaling must therefore account for climate-dependent convection-permitting dynamics rather than rely on bias-correcting current-climate operators, with direct implications for flood-risk assessment and climate-resilience planning.

Xingjian Yan, Anamitra Saha, Erin Dougherty et al. Convection-permitting scales break climate invariance in machine-learned precipitation downscaling, 23 September 2026, PREPRINT (Version 1) available at Research Square [https://doi.org/10.21203/rs.3.rs-10899014/v1]