Efficient Flow Matching for
Climate Downscaling

Abstract

High-resolution climate information is essential for understanding local impacts of climate change, yet global models remain too coarse for regional applications. This work explores generative neural networks for climate downscaling, transforming coarse climate model outputs into fine-scale projections. We focus on Flow Matching, a recent alternative to diffusion models, to efficiently learn continuous mappings between low- and high-resolution climate fields.

Using the NGCD gridded datasets at 1 km spatial resolution and daily temporal resolution across the Nordics, the model is trained to learn spatial patterns and reconstruct fine-scale temperature fields from E-OBS gridded dataset at 10 km resolution. Flow Matching provides a stable and efficient framework for high-resolution generation while allowing incorporation of geographical features and physical knowledge.

The model is trained on NGCD daily 2-m temperature fields at 1 km resolution across the Nordic region. Coarse E-OBS fields at 10 km resolution are first bilinearly upsampled to the NGCD grid and used as the starting point for downscaling. To providde geographical and temporal context, the model also receives auxilary inputs including elevation, land-water mask and (optionally) the previous two days of low-resolution temperature fields to help preserve variable dynamics.

Downscaling starts from bilinear upsampling of the E-OBS data to the 1 km grid, after which the Flow Matching framework iteratively reconstructs the spatial details. During training, the model learns a flow field that transforms the input toward the high-resolution target along a continuous trajectory. All inputs are processed as overlapping 128 ✕ 128 patches. This patch-based design enables efficient high-resolution projections over large domains (1900 ✕ 1400 px covering three Nordic countriers) while impoving spatial robustness.

We evaluate multiple configurations, including models with and without auxiliary features and models incorporating physical consistency constraints. Models are trained on years 2000–2020 and evaluated on unseen days in 2024 using mean absolute error (MAE).

Model input, target, and prediction over southern Norway and Finland in June 2024. The Flow Matching framework is able to reconstruct fine-scale geographic detail and correct input temperatures. Differences between input and target data vary across regions and seasons as well as individual dates.