Climate models often miss local extremes like intense rainfall because their resolution is too coarse. Downscaling methods improve these coarse projections to neighborhood scale, but do they still work when the climate changes?

We tested this using two advanced methods: a statistical model called the Ens-CGP and a deep learning model called RaGAN, along with a basic interpolation method. Using climate simulations over Puerto Rico, we trained each method on present-day conditions (2001 to 2021) and future projections (2041 to 2061), then swapped them by applying models trained on present data to future scenarios and vice versa. We focused on downscaling rainfall during extreme events.

To evaluate performance, we analyzed the spatial patterns and scales of rainfall using spectral methods. We repeated each experiment 15 times with different random data splits to ensure reliable results.

Our results show that some methods lose accuracy when applied outside the climate they were trained on. This means that downscaling tools may need to be retrained as the climate evolves. Our study provides practical guidance for scientists, engineers, and planners on when to update these tools to ensure accurate flood risk assessment in a warming world.