Physics-Constrained Rectified Flow for Climate Downscaling
November 18, 2026
Physics-Constrained Rectified Flow (PC-RF) uses training penalties and post-sampling projections to ensure climate downscaling follows conservation laws. On ERA5 data, the projection reduced non-negativity and mass-conservation errors to zero.
HOW THIS AFFECTS YOU
●
researcherYou can use training penalties and projections to force flow-matching models to respect physical constraints like mass conservation.
●
healthThis could make climate-driven health modeling more reliable by ensuring physical accuracy in local weather projections.