High-resolution climate projections are essential for regional risk assessment; however, Earth System Models (ESMs) operate at scales far too coarse for local impacts. This review examines how machine learning (ML) downscaling can bridge this divide and addresses a key knowledge gap: how to achieve reliable, physically consistent downscaling under future climate change. This article synthesizes ML downscaling developments from 2010 to 2025, spanning early statistical methods to modern deep learning (e.g., convolutional neural networks (CNNs), generative adversarial networks (GANs), diffusion models, and transformers). The analysis introduces a new taxonomy of model families and frames the discussion around the “performance paradox”—the tendency for models with excellent historical skill to falter under non-stationary climate shifts. Our analysis finds that convolutional approaches efficiently capture spatial structure but tend to smooth out extremes, whereas generative models better reproduce high-intensity events at the cost of greater complexity. The study also highlights emerging solutions like physics-informed models and improved uncertainty quantification to tackle persistent issues of physical consistency and trust. Finally, the synthesis outlines a practical roadmap for operational ML downscaling, emphasizing standardized evaluation, out-of-distribution stress tests, and hybrid physics–ML approaches to bolster confidence in future projections.

High‐resolution climate information is essential for risk assessment and adaptation, yet the gap between coarse Earth system model output and local scales persists. We synthesize machine‐learning (ML) approaches for climate downscaling from 2010--2025 across classical methods, convolutional super‐resolution, generative models (GANs/VAE--GANs), diffusion, and transformers. We highlight what each class actually delivers for practitioners—improvements in spatial structure, calibration, and depiction of extremes—alongside limitations that remain: sensitivity to training losses and data, non‐stationarity under warming, physical (in)consistency, and reliable uncertainty quantification. We connect methodological choices (e.g., residual vs.\ plain CNNs; intensity‐aware losses; spectra‐aware evaluation; ensemble generation) to changes in verified skill and failure modes. Our assessment yields practical guidance: pair strong linear/bias‐correction baselines with structure‐ and tail‐aware metrics; stress‐test under warming; prefer probabilistic generators when ensembles are required; and evaluate multivariate coherence when multiple variables are downscaled. We close with priorities for the next decade: physics‐aware objectives, robust out-of distribution (OOD) detection and adaptation, scalable transfer across regions and resolutions, and trustworthy evaluation protocols.

Explainable AI (XAI) is critical for building trust in complex machine learning models, yet mainstream attribution methods often provide an incomplete, static picture of a model's final state. By collapsing a feature's role into a single score, they are confounded by non-linearity and interactions. To address this, we introduce Feature-Function Curvature Analysis (FFCA), a novel framework that analyzes the geometry of a model's learned function. FFCA produces a 4-dimensional signature for each feature, quantifying its: (1) Impact, (2) Volatility, (3) Non-linearity, and (4) Interaction. Crucially, we extend this framework into Dynamic Archetype Analysis, which tracks the evolution of these signatures throughout the training process. This temporal view moves beyond explaining what a model learned to revealing how it learns. We provide the first direct, empirical evidence of hierarchical learning, showing that models consistently learn simple linear effects before complex interactions. Furthermore, this dynamic analysis provides novel, practical diagnostics for identifying insufficient model capacity and predicting the onset of overfitting. Our comprehensive experiments demonstrate that FFCA, through its static and dynamic components, provides the essential geometric context that transforms model explanation from simple quantification to a nuanced, trustworthy analysis of the entire learning process.