I started a few years ago to investigate the interplay between machine learning and physics mainly for the modeling of spatio-temporal dynamics. I am collaborating on this general topic with colleagues in geophysics, numerical analysis and fluid dynamics.
I was awarded an AI Chair in 2019 – a national program from France AI strategy. The project is entitled « DL4CLIM: Deep Learning for Physical Processes ». The program will run until 2026. It is managed by french ANR, you will find the announcement page on the ANR site here, a description of the chair program is available here (to come).
I collaborate with colleagues from INRIA Paris (INRIA/Ange) and Sophia (INRIA/Epione) to the DeepNum project, focused on the numerical analysis of ML solveurs. I also collaborate with Paola Cinella from d’Alembert lab. on the recently launched LearnFluidS project: Machine-LEARNing for FLUID Simulations.
We have been mainly exploring the following issues:
- How to incorporate prior physical knowledge in machine learning models
How to leverage both prior physical background on a phenomenon and information extracted from data? How can ML models come as a complement to physical ones?
- Domain generalization
When physical models are valid on whole domains, e.g. whole oceans, ML models do not generalize outside their training domain. When trained on physical data they do not capture the underlying laws of nature and do not generalize to other contexts (space or time). This is a major limitation of data based approaches to physics.
Recent publications related to the topic of the chair at MLIA
- Conferences – to come – you can have a look at DPLP
- PhD manuscripts – Theses
- Arthur Pajot -2019- Incorporating physical knowledge into deep neural network
- Emmanuel de Bezenac – 2021- Modeling Physical Processes with Deep Learning: A Dynamical Systems Approach
- Jérémie Donà – 2022- Statistical Learning of Physical Dynamics
- Ibrahim Ayed – 2022 –Neural Models for Learning Real World Dynamics and the Neural Dynamics of Learning