Physics-Informed Machine Learning of Cloud Processes for Next Generation Earth System Models

February 26, 2026

Kara Lamb

Hosted by Jim Hurrell

Abstract

Earth System Models (ESM) encode our knowledge about the physical world, enabling both short-term weather and long-term climate prediction. Because these models cannot explicitly resolve all relevant physical processes, they rely on simplified parameterizations of sub-grid-scale processes such as clouds, turbulence, and convection, which remain a persistent source of uncertainty. While recent advances in machine learning (ML) have enabled impressive emulation of model components and even full model surrogates, physically grounded parameterizations remain essential: they provide interpretability, scientific insight, and robustness for out-of-distribution climate prediction.

Clouds remain a leading source of uncertainty in future climate projections due to their complex, multiscale nature. Recent advances in physics-informed machine learning and generative AI provide new opportunities to better connect laboratory measurements, field observations, and high-resolution simulations with the parameterizations used in global models. I will present several recent studies that applied physics-informed machine learning to discover new interpretable physical relationships from in situ observations and high-fidelity simulations to improve predictive models for cloud processes. First, I will demonstrate how neural ordinary differential equations combined with symbolic regression can learn interpretable and generalizable ice growth laws from sparse laboratory measurements. Second, I will show how data-driven reduced-order modeling can derive simplified bulk microphysics schemes from high-fidelity atmospheric simulations in an unsupervised manner. Third, I will show recent work using generative AI to infer probabilistic atmospheric histories for individual ice crystals observed during airborne field campaigns, providing new observational constraints on ice formation pathways. Finally, I will discuss how machine learning can be used to improve cloud representation at the spatial scales needed to accurately predict processes at the climate scale.

I will conclude by outlining a pathway toward machine-learning-enabled workflows that integrate discovery, development, tuning, and evaluation of parameterizations within full Earth System Models. This framework promises to accelerate the model development cycle while preserving the physical interpretability and robustness required for predictive Earth system science.