Numerical Earth system models are fundamentally limited by multi-scale error growth, unresolved physics, and computational cost. Among these limitations, inadequate representations of aerosol and cloud microphysics remain a major source of predictive uncertainty. In this talk, we explore how modern artificial intelligence (AI) methods can help address these challenges. At the process level, we replace overly simplified parameterizations of aerosol optics, droplet nucleation, and warm-rain processes with fast, accurate, data-driven surrogates that still respect basic physical constraints. At the mechanistic level, we use observational data to investigate rapid cloud adjustments to aerosol perturbations, disentangling aerosol–cloud causal relationships from their co-variability with meteorology. We highlight the role of large, AI-ready libraries derived from observations and from high-resolution cloud- and convection-resolving simulations as training and validation datasets, with particular emphasis on sampling across regimes. Using a hybrid Earth system modeling framework, we demonstrate how changes in aerosol and cloud microphysical representations propagate to top-of-atmosphere radiative fluxes, focusing on implications for consistency with observational constraints and for reducing uncertainty in Earth system predictions. We conclude with a discussion of emerging best practices for incorporating AI into Earth system modeling and prediction and for bringing Earth system science to AI, with the goal of accelerating scientific discovery and improving predictability.
Copyright. All Rights Reserved.