Downscaling Precipitation Projections using Generative AI: Benchmarking against the WRF Dynamical Climate Model

High-resolution climate projections are essential for hydrological and meteorological impact assessments, yet dynamical numerical simulations remain computationally prohibitive for large ensembles and domains. Generative AI, specifically Probabilistic Diffusion Models (DMs), offer a promising, computationally efficient alternative. Recently, these models have demonstrated skill in reproducing historical data and serving as efficient emulators of dynamical models. The question is, therefore, whether models trained on historical observations can infer the non-stationary statistics of future climate projections.

In this work, we downscale CESM2-LENS simulations over large domains using a DM trained on reanalysis data. We investigate the model's capability to bridge the scale gap between GCM outputs (~100 km resolution) and data requirements for local hydrological impact modelling (~10 km resolution) under both historical and end-of-century scenarios. Furthermore, we compare the diffusion-based approach with the outputs of the state-of-the-art WRF dynamical model, with a focus on the changes to key hydrometeorological indices. By benchmarking DM-downscaled data against both dynamically-downscaled data and GCM baselines, we aim to assess the trade-offs between computational efficiency and physical consistency, offering insights into the generalization limits of generative AI for climate change impact studies.