Intercomparison of Convection-Permitting and Machine-Learning Downscaling for Future Climate Projections over Atlantic Island Regions

High-resolution climate information is essential for robust climate change impact assessments, particularly in insular      regions where strong land–sea contrasts, steep orography, and mesoscale processes dominate local climate variability. This study presents a systematic intercomparison of convection-permitting regional climate simulations (CPMs) and machine-learning (ML) downscaling approaches for future climate projections over the Azores and Madeira archipelagos, using CMIP6 projections      as large-scale boundary conditions.

The dynamical component consists of kilometre-scale (3 km) WRF simulations driven by ERA5 and CMIP6 EC-Earth3-Veg      outputs. These CPMs explicitly resolve deep convection and mesoscale circulation, enabling a physically consistent representation of precipitation, temperature, wind, and associated extremes. Model performance is evaluated against station observations, demonstrating substantial added value relative to the driving GCMs, particularly for precipitation variability, extreme rainfall, and coastal–orographic gradients. Future projections point to a warming of around 5ºC in the SSP5 scenario by the end of the century in both regions, with Madeira losing 10% of the annual precipitation while Azores should gain around 10%. In parallel, ML-based downscaling models trained on multi-model CMIP6 ensembles and local observations are used to generate high-resolution projections for the same regions and scenarios. These approaches efficiently reproduce mean climate signals and large-scale spatial patterns, allowing the exploration of a broader range of model uncertainty at a fraction of the computational cost. The intercomparison reveals clear methodological contrasts. CPMs provide physically consistent representations of local processes and extremes but are affected by substantial local biases, whereas ML approaches strongly reduce systematic errors, while making physical interpretability more challenging. Conversely, ML downscaling offers strong advantages in ensemble size, scenario coverage, and computational scalability. Overall, the results highlight that CPMs and ML-based approaches should not be viewed as interchangeable. Instead, their differing strengths imply distinct roles in future climate projection workflows, with CPMs remaining essential for process-based and extremes-focused studies, and ML methods offering complementary value for uncertainty assessment and rapid scenario analysis in climate services and adaptation planning.

 

This work is supported by FCT, I.P./MCTES through national funds (PIDDAC): LA/P/0068/2020 - https://doi.org/10.54499/LA/P/0068/2020, UID/50019/2025, https://doi.org /10.54499/UID/PRR/50019/2025, UID/PRR2/50019/2025. The authors would like also to acknowledge the project “Elaboração do Plano Municipal de Ação Climática de Barcelos (PMACB).