Flow Matching for Convective-Scale Precipitation Downscaling
This work demonstrates that flow matching is a competitive generative framework for convective-scale precipitation downscaling, offering improved spatial skill over diffusion models.
Flow matching outperforms diffusion models for downscaling daily precipitation from 8 km to 2 km over Singapore, achieving higher fractions skill score at all thresholds and better spatial structure, but underestimates extreme precipitation leading to a dry bias.
Generative machine learning is an increasingly important complement to dynamical downscaling for producing high-resolution precipitation projections, with diffusion models currently the leading approach. Flow matching is a related generative framework that has recently achieved strong results across image, video and other domains, and shown early promise for downscaling. We train a flow matching model to map daily precipitation from 8 km to 2 km over a convective-scale domain centred on Singapore, and benchmark it against CPMGEM, a score-based diffusion model. Flow matching achieves consistently better spatial skill: higher fractions skill score at every precipitation threshold and neighbourhood scale tested, and tighter structure and amplitude components of the SAL score with comparable location skill. However, flow matching underestimates the upper tail of the precipitation distribution, resulting in a dry bias in the climatological mean. These results suggest that flow matching is a competitive generative framework for convective-scale precipitation downscaling, particularly well suited to capturing spatial structure.