Title: Applying Flow Matching to Downscale Convective-Scale Precipitation

Generative machine learning has emerged as a vital supplement to dynamical downscaling for generating high-resolution precipitation projections, with diffusion models currently holding the top spot. However, flow matching—a related generative framework that has recently delivered impressive results in image, video, and other fields—has also shown early potential for downscaling tasks. In this study, we developed a flow matching model designed to upscale daily precipitation data from an 8 km resolution to a 2 km resolution across a convective-scale domain centered on Singapore. We evaluated its performance by benchmarking it against CPMGEM, a score-based diffusion model.

The analysis reveals that flow matching delivers superior spatial skill across the board. It achieved higher fractions skill scores at every tested precipitation threshold and neighborhood scale, while also demonstrating tighter structure and amplitude components within the SAL score, maintaining comparable location skill. Despite these spatial advantages, the model exhibits a tendency to underestimate the upper tail of the precipitation distribution, leading to a dry bias in the climatological mean. These findings indicate that flow matching is a competitive generative framework for convective-scale precipitation downscaling, particularly excelling at capturing spatial structures.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC