Kilometer-scale climate data are crucial for water supply planning, wildfire risk management, and infrastructure design in the mountainous western United States, but running a regional model at this resolution for every CMIP6 Earth System Model (ESM) is prohibitively expensive. We developed a lighter, data-driven alternative: a two-stream UNet-GAN. One stream learns how variables such as precipitation and temperature interact, while the other captures each variable’s spatial patterns; their combination reduces unwanted artifacts in the downscaled data.

The network is trained on a 43-member WRF ensemble spanning 1980–2100. Even when using only one-third of all days, it lowers grid-cell root-mean-square error by about 26% for rainfall and 48% for temperature compared with bilinear interpolation, while also outperforming strong single-stream machine learning baselines (e.g., UNet) across both fields. Adding more data helps initially, but gains approach saturation. Training on historical data alone misses future warming and precipitation trends due to non-stationarity, so we mix in future samples to preserve long-term trends and extremes.

The model reproduces key physical patterns - precipitation maxima align with orographic crests in both climatological means and extremes - and retains skill when applied to entirely different ESMs, showing it learns transferable physics rather than memorizing one dataset.