Data-driven discovery of fluid closures for magnetic reconnection from fully kinetic particle-in-cell simulations

Accurate fluid descriptions of collisionless plasmas are important to describe and understand the dynamics of magnetic reconnection in large-scale systems. However, one of the main challenges of such fluid descriptions is the development of accurate closure models that capture the important effects of unresolved higher-order moments of the distribution function on the lower-order moments. In this work, we explore the use of machine learning techniques to learn accurate closure models for the plasma pressure and heat flux tensors from the data of fully kinetic simulations of magnetic reconnection. We show the importance of embedding fundamental physical symmetries such as frame-invariance into the machine learned models to eliminate spurious correlations/dependencies in the data and to improve generalizability to different system sizes and plasma conditions. We also investigate the role of non-locality on the accuracy of the closure models using convolutional neural networks, and determine an optimal kernel sizes that balances between closure accuracy and complexity. Finally, we discuss the possibility of integrating these machine-learned closures into the loop of fluid plasma simulations.