Machine learning and numerical models in geoscience

Séminaire de J. Brajard (LOCEAN) et R. Lguensat (IGE), Lundi 9 Juillet 2018 à 14h en salle Lliboutry, Bât. Glaicologie

Séminaire "Machine Learning in Geoscience" :

******** Présentation 1 : Julien Brajard, Equipe LOCEAN, UPMC

Title : Machine learning and numerical models in geoscience.

In collaboration with : Anastase Charantonis, Jérôme Sirven, Emmanuel de Bozenac, Arthur Pajot and Patrick Gallinari.

Abstract : Numerical models are used to simulate the evolution of atmosphere or ocean dynamics. They are implemented through a computer code, that contains predefined rules specifying how to compute the evolution of some outputs (e.g sea surface height) from inputs (e.g. previous states of the model, satellite or in situ observations of other parameters). A machine learning approach, in contrast, infers its internal set of rules from a large amount of data. In many fields (image recognition, automatic translation, speech recognition, ...), the more traditional methods, which rely on predefined rules, have been outperformed by machine learning algorithms. This performance was made possible by advances in Convolutional and Recurrent Neural Networks.

This work addresses the question of the application and the usefulness of machine learning for numerical modeling in Geophysics. A first result presents the capacity of neural networks to correct ill posed parametrization operators in a idealized shallow-water model. Another application to a fully data-driven model is presented, and finally some preliminary results are shown on data assimilation using machine learning.

******** Présentation 2 : Redouane Lguensat, Equipe MEOM, IGE

Title : Machine Learning for the inversion of SWOT-Ocean data

Abstract : The future SWOT (Surface Water Ocean Topography) space mission offers an unprecedented opportunity for observing oceanic small-scale processes and for assessing their impact on global ocean circulation and climate. Observations of sea surface height (SSH) at scale ≤ 100km could indeed allow to better understand energy cascades toward dissipative scales in the ocean, and to better understand the processes involved in vertical exchanges of heat and tracers between the ocean surface and the ocean interior.

Over the last decade, impressive advancements have been made in machine learning (ML) methods, mostly thanks to the revival of neural networks and the large increase in computational power. This has led to the birth of "deep learning" which is now ubiquitous in signal and image processing applications. This work aims to investigate the use of one class of DL techniques, namely, Convolutional Neural Networks (CNNs) in the context of SWOT-Ocean related inverse problems. We highlight three selected inverse problems that can be tackled by the use of CNNs and show first results. We then discuss challenges and opportunities that come with applying machine learning in ocean data in general.