Applications for a CEA-funded PhD, on the topic "Potential and pitfalls of using machine learning in climate modelling", are open.
If interested, please send your CV, academic transcripts, letter of motivation and name of contacts for recommendations to:
Masa Kageyama (LSCE, Masa.Kageyama _at_ lsce.ipsl.fr), Mathieu Vrac (Mathieu.Vrac _at_ lsce.ipsl.fr) and Thomas Dubos (LMD, Thomas.Dubos _at_ lmd.ipsl.fr).
Anticipating future climate change is vital for our societies. Numerical climate modelling plays an essential role for this anticipation. Climate models are first designed to represent recent climate. Showing their capacity to represent past cold or warm climates is also essential for increasing their credibility in representing climates different from present, never experienced by humankind. Climate modelling is based on fundamental physical principles, which are explicitely resolved in the case of phenomena of large-enough spatio-temporal scales, or represented statistically or semi-expirically for smaller scale phenomena. Machine learning is a ‘physics-free’ statistical modelling technique which has recently greatly improved in various domains. Pioneering work [Gentine et al, 2018] suggests that part of a climate model, in particular sub-grid-scale parameterizations, could be represented via results of analyses based on machine learning and hence help the models to gain computational efficiency. The objectives of this PhD is to explore the potential and pitfalls in building such new models partly based on machine learning techniques, for representing present, past and future climates. The project will make use of the LMDZ atmospheric general circulation model. The work will first examine idealized configurations (aquaplanets, no land) and then progress towards more realistic protocols, first including the coupling to the land surface, and if possible coupling to the ocean.
Supervisors: Masa Kageyama, Mathieu Vrac (LSCE) and Thomas Dubos (LMD). This supervision team will be completed by Yann Meurdesoif and Sébastien Fromang, on technical aspects related to the ico-LMDz model, and the ESTIMR team for the machine learning aspects.