Machine learning (ML) is revolutionizing atmospheric modeling across scales, yet ML models may violate physical laws, struggle outside their training set, and explaining their added value remains challenging—especially for deep learning models. This presentation explores a two-way synergy between ML and physical knowledge: (1) using physics to constrain or guide ML to improve its consistency and generalizability across atmospheric regimes, and (2) distilling knowledge from successful ML models via Pareto-optimal model hierarchies. I will demonstrate this with case studies, including improving the generalization of neural network parameterizations across climates, discovering equations linking cloud cover to its thermodynamic environment, and elucidating three-dimensional patterns in radiative feedbacks associated with early tropical cyclone intensification. While the focus is on weather and climate applications, the methodological frameworks apply broadly to scientific ML, with the dual purpose of improving the trustworthiness of ML for environmental applications and facilitating data-driven discovery in Earth sciences.

Tom Beucler, Univ. Lausanne.

Lieu
École normale supérieure – PSL
24 rue Lhomond – aile Erasme
salle Lien Hua – E350