PIF2025 - Advancing machine learning models for downscaling and their evaluation - (PID2024-162703OB-I00)

This thesis advances machine learning approaches for regional climate downscaling and their systematic evaluation, addressing long-standing limitations of traditional statistical methods and computationally expensive Regional Climate Models (RCMs). Classical empirical–statistical downscaling has been applied mainly at local scales and cannot be easily extended to continental domains, limiting its contribution to coordinated CORDEX experiments. Building on the foundations laid by the ATLAS project, recent advances demonstrate that deep learning models can learn complex spatiotemporal relationships and outperform conventional methods, enabling more scalable downscaling strategies. These include super-resolution techniques, perfect-prognosis downscaling, and emerging RCM emulators capable of reproducing regional climate dynamics. This thesis contributes to developing the next generation of ML-based downscaling methods suitable for continental-scale application and supporting CORDEX activities by complementing traditional RCM ensembles. A second key component of the work is the design of process-based evaluation metrics to assess model realism, transferability, and extrapolation skill, moving beyond error statistics to evaluate whether models capture the dynamical mechanisms driving regional climate. These evaluation tools align with ongoing CORDEX and CMIP7 efforts to establish robust, globally applicable criteria for model selection and performance assessment.