Advances in aggregation methods for multi-model ensembles in climate science: A comprehensive review

Multi-model ensembles (MMEs) have become a cornerstone of climate modeling, offering a structured approach to addressing uncertainty and variability in climate projections. This review critically synthesizes the evolution of MME methodologies, including statistical, probabilistic, machine learning (ML), and hybrid approaches, highlighting key advances, persistent limitations, and future research opportunities. Early MME methods, such as simple averaging and performance based weighting, provided foundational insights but were unable to capture inter-model dependencies and complex nonlinear dynamics. Reliability Ensemble Averaging (REA) and Bayesian Model Averaging (BMA) enhanced projection skill through probabilistic frameworks, yet faced challenges related to natural variability and prior specification. Recent advances in ML and hybrid techniques now enable more flexible, accurate, and robust ensemble aggregation across diverse climatic contexts. Despite significant progress in developing statistically rigorous and application specific ensemble frameworks, challenges remain across several dimensions, including model interdependence, structural uncertainty, static weighting schemes, computational demands, and limited usability for non-expert stakeholders. Emerging solutions, such as skill and independence based weighting, hybrid and ML approaches, and deep generative models, are advancing ensemble diversity, regional relevance, and physical plausibility. This review outlines a forward looking research agenda that emphasizes modular, interpretable, and scalable ensemble systems, time varying weighting strategies, robust stress testing protocols, and co-designed, sector specific applications. Addressing these gaps is essential to enhance the credibility, usability, and decision making relevance of MMEs in the face of a rapidly changing climate.