Multidisciplinary Institute in Artificial intelligence

The goal of SBI4C is to develop ML methods to better understand and model physical phenomena, with a specific interest in climate sciences and inverse problems that appear in this context. Simulation-based inference (SBI) enables the estimation of non-linear model parameters from observational data, which are useful, for example, when tuning the parametrizations of atmospheric and oceanic models. However, current SBI methodology has been limited to small-scale models and our project seeks to extend this framework. We will address two main challenges: craft new SBI algorithms that take the computational expense of simulations into account and propose strategies to generate data based on efficient emulators. We expect methodological and software contributions, applications to climate models with experiments on supercomputers, and transfers to industrial partners.