Data-driven modeling of collisional kinetic plasma dynamics

Computational plasma physics has seen significant advances in the ability to model nonlinear plasma dynamics from first-principles. However, capturing the complex multi-scale dynamics of plasmas, especially in non-equilibrium and strongly coupled regimes, remains challenging. This talk will highlight recent work that explores the potential of combining machine learning algorithms with traditional plasma kinetic simulation frameworks to address these challenges.


The first part of the talk will focus on integrating  Graph Neural Networks (GNNs) as surrogate models for kinetic plasma simulators.  We demonstrate that GNNs can accurately replicate the dynamics of a one-dimensional plasma for arbitrary degrees of collisionality, and generalize effectively to conditions well beyond their training data. We also show that GNNs can implicitly learn a computationally efficient collisional operator in 1D.

In the second part, we discuss the use of self-consistent electromagnetic Particle-in-Cell (PIC) simulations that resolve the inter particle fields that mediate collisional interactions to guide the development of data-driven, collisional operators. These data-driven operators can shed light on the nature of collisions in strongly coupled regimes and, in the future, elucidate how collisional operators are modified in far-from-thermodynamic-equilibrium conditions.