DefinePK

Abstract

Climate modeling and extreme weather forecasting are at the center of efforts to understand and respond to global climate change. Traditional numerical models remain the scientific backbone of prediction, but they are limited by computational expense, uncertainties in representing small-scale processes, and difficulty capturing extremes. Recent advances in machine learning (ML) have opened new possibilities to complement or partially replace physics-based models. ML techniques provide powerful tools for learning nonlinear relationships, correcting systematic biases, downscaling coarse climate simulations, and detecting rare events. This paper reviews the state of research on ML applications in climate modeling and extreme weather forecasting. We organize the literature around surrogate modeling, parameterization of sub-grid processes, emulators, post-processing and bias correction, downscaling, and hybrid physics-ML methods. We then survey ML applications in forecasting rare and extreme events such as tropical cyclones, floods, droughts, and heatwaves. To strengthen the review, we propose a hypothetical experiment on ML-enhanced extreme precipitation forecasting, outlining data sources, architectures, and evaluation metrics. The paper concludes with a discussion of methodological challenges, including interpretability, generalization under climate change, uncertainty quantification, and deployment in operational systems. We argue that hybrid approaches combining physical knowledge with data-driven models represent the most promising pathway forward.