Embedding physical symmetries into machine-learned reduced plasma physics models via data augmentation

Abstract: Machine learning is offering powerful new tools for the development and discovery of reduced models of nonlinear, multiscale plasma dynamics from the data of first-principles kinetic simulations. However, ensuring the physical consistency of such models requires embedding fundamental symmetries of plasma dynamics. In this work, we explore a symmetry-embedding strategy based on data augmentation, where symmetry-preserving transformations (e.g., Lorentz and Galilean boosts) are applied to simulation data. Using both sparse regression and neural networks, we show that models trained on symmetry-augmented data more accurately infer the plasma fluid equations and pressure tensor closures from fully kinetic particle-in-cell simulations of magnetic reconnection. We show that this approach suppresses spurious inertial-frame-dependent correlations between dynamical variables, improves data efficiency, and significantly outperforms models trained without symmetry-augmented data, as well as commonly used theoretical pressure closure models. Our results establish symmetry-based data augmentation as a broadly applicable method for incorporating physical structure into machine-learned reduced plasma models.