A talk by Tom Beucler, Assistant Professor of Environmental Data Science at the University of Lausanne, Switzerland, hosted by Leeds Institute for Data Analytics' (LIDA) Scientific Machine Learning (SciML) group.
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.
Key references:
1. Beucler, T., Grundner, A., Shamekh, S., Ukkonen, P., Chantry, M., & Lagerquist, R. (2024). Distilling machine learning's added value: Pareto fronts in atmospheric applications.
arXiv e-prints, arXiv:2408.
2. Beucler, T., Gentine, P., Yuval, J., Gupta, A., Peng, L., Lin, J., … & Pritchard, M. (2024). Climate-invariant machine learning. Science Advances, 10(6), eadj7250.
3. Iat-Hin Tam, F., Beucler, T., & Ruppert, J. H., Jr. (2024). Identifying three-dimensional radiative patterns associated with early tropical cyclone intensification. Journal of Advances in Modeling Earth Systems, 16, e2024MS004401.
4. Beucler, T., Pritchard, M., Rasp, S., Ott, J., Baldi, P., & Gentine, P. (2021). Enforcing analytic constraints in neural networks emulating physical systems. Physical Review Letters, 126(9), 098302.
5. Zanetta, F., Nerini, D., Beucler, T., & Liniger, M. A. (2023). Physics-constrained deep learning postprocessing of temperature and humidity. Artificial Intelligence for the Earth Systems, 2(4), e220089.
Bio:
Tom Beucler is an assistant professor of environmental data science at the University of Lausanne, Switzerland. In August 2021, he established the Data-Driven Atmospheric & Water Dynamics (∂3AWN) laboratory, the first research group worldwide dedicated to bridging atmospheric physics and artificial intelligence. His research group combines physical and dynamical theory, computational science, statistics, observational analyses, and numerical simulations to enhance our understanding of the atmosphere while improving weather forecasting and climate predictions. Before joining the University of Lausanne, Tom earned his Ph.D. in atmospheric science from MIT, where he studied the interaction between water vapor, radiation, and convection in the tropics. He then pursued postdoctoral research at Columbia University before becoming a project scientist at the University of California, Irvine, focusing on integrating physical knowledge into neural network representations of cloud processes for climate modeling.
Group website: https://wp.unil.ch/dawn/
GitHub: https://github.com/tbeucler
Google Scholar: https://scholar.google.com/citations?user=OODC5e8AAAAJ&hl=en
ResearchGate: https://www.researchgate.net/profile/Tom-Beucler
LinkedIn: https://www.linkedin.com/in/tom-beucler/
Email address: [email protected]