Deep Learning for Physical Processes with applications to Earth System Science – DL4CLIM

Motivations and Scientific Program: the project targets the development of Deep Learning (DL) methods for the modeling of physical processes. The application domain is environment and climate. It builds on the complementarity of two major scientific paradigms, Physics and Machine Learning (ML). The former relies on elaborate and complex models of natural phenomena but does not offer principled methods for integrating the data generated by observation platforms (e.g. satellite) and climate models. The latter develops an agnostic data centered approach but faces major challenges for modeling complex physical phenomena. Our objective is to answer these challenges by developing modeling systems coupling knowledge based physical process models with data-driven machine learning. We believe that this is a major scientific challenge for the upcoming years and that its impact can be much bigger than what has been achieved recently in engineering domains like computer vision.
The project will focus on the modeling of spatio-temporal processes characteristic of the environment and climate dynamics. These processes are governed by general laws usually modeled by partial differential equations characterizing fluid dynamics. The objective is then to develop hybrid systems able to learn these dynamics from data. It is organized in two main tracks: fundamental ML developments and use cases in environment, both developed in close interaction. The fundamental aspects cover two topics. The first one is the development of hybrid physical-statistical systems and proceeds in three successive steps: (1) learning of known PDEs from simulated data, (2) learning unknown dynamics from incomplete observations and (3) incorporation of physical priors in Deep Learning models. Creating hybrid models is only one part of the problem and developing Machine Learning for Earth System Science requires solving specific learning problems. For the second topic we consider typical ML problems motivated by our use cases.
For the second track, three use cases have been selected illustrating a variety of representative Earth System Science (ES) problems. They respectively concern: (1) the improvement of ocean current circulation models by integrating high and low resolution satellite information, (2) the detection and tracking of Eddies which are known to have a strong impact on the biological productivity of the ocean, (3) a more prospective topic: modeling the influence of anthropic forcing (Greenhouse gases, Ozone, etc.) on climate change.

Team: the PI leads a pluri-disciplinary team composed of 3 ML and 4 ES specialists, all working in close collaboration. The participants have already collaborated through a pluri-disciplinary working group launched 2 years ago at Sorbonne and through joint mentoring of internships. 2 invited professors will join the project.

Impact: Sorbonne has launched a center for AI in 2019 (SCAI) aimed at promoting core AI and cross-disciplinary research. Environment is one among the 3 cross-disciplinary axis selected for SCAI. The PI is co-responsible of this SCAI axis together with an ES colleague. At the national level, the project addresses 4 out of 6 of the main directions highlighted in the 2018 report by French deputy Villani, which has been the basis for the French strategy on AI. Environment is one of the 4 priority application fields identified in this strategy. The PI is heavily involved in the development of ML teaching and dissemination activities at Sorbonne through the Computer Science and the Mathematics master curriculum. He will be responsible for a new joint Computer Science --Mathematics Data Science master program to start next year (2019-2020). He is also involved in lifelong learning and in the organization of special sessions on Machine Learning and Deep Learning. The project will be the opportunity to enlarge the training programs and to reach a pluri-disciplinary audience, from master students to researchers.