Improving Interpretable Machine Learning for Plasmas: Towards Physical Insight, Data-Driven Models, and Optimal Sensing

Magnetized plasmas, a combination of superheated gas and magnetic fields, are pervasive in our universe and are responsible for some of the grandest natural phenomena, such as the aurora. Plasmas are also extensively studied for engineering and industrial applications, such as space propulsion and development of future fusion energy reactors. This project aims to improve our ability to understand and predict the behavior of magnetized plasmas using simplified models that are both fast and easy to use. In particular, this investigation will explore methods that combine machine learning techniques that are revolutionizing many fields, like self-driving cars, with the known physical laws that govern magnetized plasmas - seeking to leverage the best aspects of each individual approach.This project will advance data-driven modeling approaches such as machine learning by utilizing physics-informed constraints for magnetized plasmas in three ways: 1) Several emerging data decomposition methods will be applied to numerical simulations of magnetized plasmas for the first time and assessed for these systems; 2) Data-driven nonlinear models based on these decompositions will be tested for modeling magnetized plasmas with significantly increased speed compared to classical approaches; 3) Methods to optimize the placement of sensors to diagnose magnetized plasmas will be evaluated to improve the value of measurements used to both observe plasmas and as the source of information to build data-driven models. Together these three studies will advance the effectiveness of low-dimensional, nonlinear, and interpretable data-driven methods for achieving new physical insight, improved prediction, and robust control of multi-scale magnetized plasmas.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.