Potential for Machine Learning Emulators to Augment Regional Climate Simulations in Provision of Local Climate Change Information

Abstract High-resolution regional climate simulations provide detailed information on future climate change to support decision-making. Ensembles of simulations, including at kilometer-scale resolution, are becoming available from international coordinated initiatives, but these do not effectively sample the full range of uncertainties. Machine learning (ML) has already been used for statistical downscaling but has the potential to augment high-resolution simulations, via emulators, enabling rapid production of local climate information at a fraction of the cost. Here, we explore skill in ML-based emulators sampling a range of architectures and identify remaining scientific issues that need to be addressed before such emulators can be considered ready for application in climate services. This includes the ability to capture extremes, produce coherent multivariate predictions, account for memory in the climate system, and robustly downscale other (out of sample) global climate models. Climate expertise needs to be integrated into the development and evaluation of ML emulators, and here, we provide recommendations on validation methods. If skillful, ML emulation has implications for how we coordinate and perform regional climate simulations. We should focus on running at the highest resolution and greatest Earth system complexity affordable, to give the best representation of processes at the local scale, for subsequent training of ML emulators. Emphasis should be on sampling the full range of conditions, including high-end scenarios. Overall, ML has promise to augment our production of regional-to-local climate projection information over the next 5–10 years, and as a climate community, we need to come together to address the relevant scientific issues. Significance Statement This study explores the potential for machine learning (ML) emulators to enhance our production of local climate change information. This is important because high-resolution regional climate projections, needed to support local decision-making for climate change adaptation, are costly and will never be able to sample the full range of uncertainties. The paper identifies the remaining open science issues and priority research directions for making these ML techniques truly useful. It also provides recommendations for specific evaluation methods, going beyond what has been done in ML studies to date. This aims to bring climate and weather expertise into the development and evaluation of ML-based emulators, disentangling the good and the bad of this new hybrid downscaling family.