Discovering novel physics using differentiable kinetic simulations


In this work, we show a proof of concept that optimization of plasma physics simulations using gradient-based methods can work. We also show that by applying the correct loss function, one can tease out previously unknown physics. This is an example of AI-based physics discovery. The skeptic might say that this is just gradient-descent based physics discovery. A moderate and fair take is that AI-based and gradient-descent-based (given the nature of deep learning anyway), is not so different in 2023. Perhaps it is better to say deep-learning based physics discovery. That is more defensible!

This article was listed as a featured article in the Journal of Plasma Physics for most of 2023!

[Link to JPP article]