CDS Faculty Advance Climate Science through Machine Learning with M²LInES
The international research collaboration works to innovate current climate models
Machine learning and data science are used in the field of climate science to advance research and solutions to the global crisis. One innovative collaboration is Multiscale Machine Learning In Coupled Earth System Modeling (M²LInES) which seeks to expand the current understanding of the climate process, create new “physics-aware” machine learning models that can advance climate research, and improve existing climate models. The program is specifically focused on research at the “air-sea-ice interface.”
The initiative brings together a team of international researchers from Princeton, Columbia, MIT, the Lamont-Doherty Earth Observatory (LDEO), the National Center for Atmospheric Research (NCAR), the French National Centre for Scientific Research (CNRS-IGE) along with Institut Pierre-Simon Laplace (CNRS-IPSL), and New York University. CDS faculty involved in the project are Joan Bruna and Carlos Fernandez-Granda along with CDS affiliated professor Laure Zanna. “The M²LInES collaboration has been extremely interesting and challenging,” said Carlos. “We are trying to apply ideas from machine learning to address real problems in climate science.”
Joan is an Associate Professor of Computer Science and Data Science who works with the CILVR (Computational Intelligence, Learning, Vision, and Robotics) group at NYU. His research interests cover machine learning, signal processing, and high-dimensional statistics. As an Associate Professor of Mathematics and Data Science, Carlos researches the design and analysis of data science methodology. The focus of his group is on machine learning applications to medicine, climate science, and scientific imaging.
Laure Zanna is a Professor in Mathematics and Atmosphere/Ocean Science who researches climate system dynamics, specifically studying the influence of the ocean on local and global scales. At NYU, she heads the Climate and Ocean Physics group that explores our understanding of ocean dynamics to improve climate change projections. She is currently the lead principal investigator for M²LInES as well as for the NSF-NOAA Climate Process Team on Ocean Transport and Eddy Energy, and the Geoscience Director for the NSF Science and Technology Center at LEAP (Learning the Earth with Artificial Intelligence and Physics).
All three researchers have recently published papers listed on the M²LInES publication page. Joan’s work “On Gradient Descent Convergence beyond the Edge of Stability” analyzes gradient descent (also known as the steepest descent) in the regime where the step-size (its critical hyperparameter) is beyond the ideal threshold dictated by traditional theory, motivated by modern ML applications where such theory does not apply. Carlos and Laure both worked on “Benchmarking of machine learning ocean subgrid parameterizations in an idealized model” which compares the performance of machine learning models used to parameterize specific processes in current ocean models, including a new algorithm developed by the NYU team to discover symbolic expression from data. Carlos and Laure also collaborated on “Deep Probability Estimation” published in the Proceedings of the 39th International Conference on Machine Learning.
By Meryl Phair
