From climate projections to trajectory-conditioned local predictions

Abstract. Regional climate projections are usually sharpened by constraining models’ forced responses toward observations, a step that filters out internal variability and works well mainly where the signal-to-noise ratio is high. Here we side-step this limitation with deep learning, treating internal variability not as noise to remove but as part of the future to predict and of the uncertainty to report. A U-Net trained on 290 simulations from 40 CMIP6 Earth System Models maps the 1980 to 2025 evolution of surface air temperature onto decadal-mean monthly climatologies from the 2030s to the 2090s under SSP2-4.5, keeping the forced response and internal variability together. Tested on entirely withheld model families, the forecasts are skilful and empirically calibrated out-of-sample. Within a model's own ensemble, a forecast predicts the future of the specific member it was given more accurately than the futures of that model's other members, a skill that persists to the 2090s, notably over the subpolar North Atlantic, Europe, North America and Southern Ocean. The forecasts also generalise to a next-generation eddy-rich model whose resolved physics lies outside the training distribution, supporting application to real-world data. When applied to observations, the framework reduces uncertainty in global-mean warming similarly to observational constraints, while also providing spatially coherent local predictions. This shows that our deep learning approach can estimate the distribution of future local climate, conditioned on the recent observed trajectory, encompassing uncertainty arising from both forced response and internal variability realisations. This extends observationally constrained projection toward calibrated predictions at the local scale and seasonal resolution where adaptation decisions are made.