This review explores the potential of machine learning (ML) in enhancing regional climate downscaling, a process crucial for improving the accuracy of climate projections at local scales. The authors highlight the limitations of global climate models (GCMs), which have coarse spatial resolutions, and discuss two main approaches for downscaling: dynamical downscaling and empirical downscaling. Dynamical downscaling, while physically based, is computationally expensive, limiting its practical application. In contrast, empirical downscaling, which includes statistical and ML techniques, is more computationally efficient and can handle larger ensembles of GCMs. The review focuses on recent advancements in ML methods, particularly deep learning, and their application in observational downscaling and regional climate model (RCM) emulation. It also addresses the challenges of domain adaptation, the need for transparency and interpretability in ML algorithms, and the evaluation of these algorithms' performance in historical and future climates. The authors propose an evaluation framework for ML-based empirical downscaling algorithms and discuss how these advancements can contribute to collaborative initiatives like the Coupled Model Intercomparison Project (CMIP) and the Coordinated Regional Downscaling Experiment (CORDEX). Overall, the review emphasizes the transformative potential of ML in improving the accuracy and reliability of climate downscaling.This review explores the potential of machine learning (ML) in enhancing regional climate downscaling, a process crucial for improving the accuracy of climate projections at local scales. The authors highlight the limitations of global climate models (GCMs), which have coarse spatial resolutions, and discuss two main approaches for downscaling: dynamical downscaling and empirical downscaling. Dynamical downscaling, while physically based, is computationally expensive, limiting its practical application. In contrast, empirical downscaling, which includes statistical and ML techniques, is more computationally efficient and can handle larger ensembles of GCMs. The review focuses on recent advancements in ML methods, particularly deep learning, and their application in observational downscaling and regional climate model (RCM) emulation. It also addresses the challenges of domain adaptation, the need for transparency and interpretability in ML algorithms, and the evaluation of these algorithms' performance in historical and future climates. The authors propose an evaluation framework for ML-based empirical downscaling algorithms and discuss how these advancements can contribute to collaborative initiatives like the Coupled Model Intercomparison Project (CMIP) and the Coordinated Regional Downscaling Experiment (CORDEX). Overall, the review emphasizes the transformative potential of ML in improving the accuracy and reliability of climate downscaling.