Lucas Meyer

Description

Many engineering and scientific applications rely on large ensembles of numerical simulations that reproduce faithfully complex phenomena. These ensembles are limited by long computation time and finite storage capacity. These issues led the development of high performance computing (HPC) and reduced order modeling techniques. Recently, the success of deep learning led the scientific community to consider its use for accelerating numerical simulations. However, deep surrogate models require many simulations for training. This approach thus suffers the same limitation that motivates its development in first instance : to produce a representative training dataset of faithful simulations is tedious. We propose an online training framework for deep surrogate models that generate simulation data on-the-fly by leveraging HPC resources.