A machine-learning framework for spatio-temporal statistical downscaling of CMIP6 climate projections

Regional ocean simulations require high-resolution atmospheric forcing, whereas CMIP6 climate projections are typically too coarse in both space and time for direct application. In the MOIRAI project, we developed a spatio-temporal statistical downscaling framework to produce multi-decadal, high-resolution 3-hourly atmospheric forcing over the Mediterranean, North Sea, and Arctic domains. The methodology consists of two main steps. First, daily climate model fields were spatially downscaled to the target regional grids and bias-adjusted using high-resolution reanalysis datasets as reference. Second, the daily bias-adjusted fields were temporally disaggregated to 3-hourly resolution using machine-learning models trained on reanalysis data for 1985–2014. The framework was applied to key forcing variables, including 2 m temperature, relative humidity, mean sea level pressure, 10 m wind, precipitation, and radiative fluxes. A range of machine-learning methods was evaluated, including support vector machines, random forests, k-nearest neighbours, XGBoost, and neural networks. As the overall predictive skill was broadly similar among methods, XGBoost was selected for implementation because it provided the best compromise between performance, computational efficiency, and scalability for long transient climate simulations.

The results show that the framework reproduces realistic sub-daily variability and coherent spatial patterns, particularly for variables with a pronounced diurnal cycle, such as temperature, relative humidity, and radiative fluxes. The analysis also highlights clear differences in reconstruction skill among variables. Predictors at daily resolution provide stronger constraints for variables dominated by the day–night cycle, whereas sub-daily reconstruction is more challenging for variables such as mean sea level pressure, whose variability is largely driven by synoptic-scale dynamics rather than diurnal forcing. In these cases, the method preserves the large-scale daily structure, but the exact hour-to-hour evolution may differ from the reference reanalysis.

Overall, the proposed framework offers a computationally efficient approach for transforming coarse-resolution global climate projections into high-resolution atmospheric forcing suitable for regional marine and coastal climate applications.

Funding This work was supported by the European Union through the MOIRAI project, Grant Agreement No. 101180994.