Deep learning-based enhancements in computational physics

Abstract: Traditional numerical algorithms in computational physics, such as methods to solve partial differential equations or to solve inverse-problems, often suffer from bottlenecks. These may manifest in the form of computationally expensive components of the algorithm, or problem-dependent parameters that require empirical tuning.

In this talk, I will present several novel data-driven approaches to overcome such bottlenecks. In particular, I will demonstrate how deep neural networks can be trained to assist and improve numerical methods. This includes the detection and control of spurious Gibbs oscillations encountered while using high-order methods to approximate solutions with low-regularity, approximation of parameter-to-output maps in many-query problems, and the representation of high-dimensional priors in Bayesian inference.

YoutubeLink to the recording