Physics-Informed Machine Learning to Reconstruct Plasma Dynamics in Basic Plasma Science Experiments

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

Plasmas — hot, electrically charged gases that make up most of the visible universe — play a central role in many areas of science and technology, from forecasting space weather to developing revolutionary energy systems like fusion reactors. Laboratory experiments are essential for understanding how plasma behaves, but measuring it in full is extremely difficult. Plasma's high temperature and fast-changing nature mean that only a portion of what is happening can usually be captured, leaving many pieces of the picture missing. This project aims to develop new machine learning (ML) tools that are guided by the laws of physics to help complete that picture. By combining limited measurements from experiments with theoretical models, these tools aim to reconstruct hidden or hard-to-measure aspects of plasma behavior. If successful, this approach will lead to better understanding of plasma dynamics, transforming our ability to extract insight from cutting-edge experiments. The ML tools may also be applied to future multi-satellite space missions, where scattered satellite measurements need to be stitched together into a cohesive view. The project will train the next generation of researchers at the intersection of physics, computation, and machine learning. The Large Plasma Device (LAPD) at UCLA is a unique experimental platform for basic plasma science, enabling studies relevant to space physics, astrophysics, and controlled nuclear fusion. Its high reproducibility, 1 Hz repetition rate, and comprehensive suite of diagnostics allow for detailed spatiotemporal measurements of complex, nonlinear plasma dynamics. Despite these capabilities, the inherent nature of plasmas — characterized by high dimensionality and multiscale behavior — means that experimental measurements of the underlying physics are inevitably incomplete. To bridge this gap, this project aims to develop and apply novel techniques from the emerging field of physics-informed machine learning. These tools will integrate partial measurements from multiple diagnostics with theoretical plasma models to reconstruct quantities that are unmeasured or difficult to observe in dynamically evolving plasmas. The approach will first be developed and validated using nonlinear plasma dynamics data from fully-kinetic first principles simulations, serving as controlled “numerical experiments.” It will then be applied to real experimental data from the LAPD, using it as a testbed to demonstrate the methodology on Alfvén wave dynamics. This validation will serve as a critical stepping stone toward future applications of the method to more complex laboratory experiments, such as studies of collisionless shocks and magnetic reconnection. Moreover, these techniques are expected to have broader applicability to space plasma observations and fusion reactor diagnostics, where sparse measurements must be synthesized into coherent global reconstructions. 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.