DeepFakesFLY

Simulating kinetic plasma dynamics is an extremely heavy computational task. For this purpose, massively parallelized simulation codes based on the Particle-in-Cell (PIC) method have been extensively optimized over the years to take advantage of existing supercomputing infrastructures. As we study more extreme multi-scale plasma phenomena, performing global kinetic simulations can become computationally infeasible. For this reason, discovering reduced models of plasma dynamics is currently an extremely active area of research. In particular, data-driven fluid closures that can be integrated into fluid codes to simulate the learned dynamic at a significantly reduced computational cost. This project aims to learn fluid closures using neural network models and first-principles PIC simulation data. We will explore neural network architectures and data augmentation methods that preserve fundamental physical symmetries and constraints, for e.g. rotational/translational symmetry and Galilean/Lorentz invariance, and integrate non-local (spatial and temporal) information. Furthermore, we will compare our results against existing theoretical and competing data-driven approaches. The requested GPU resources will enable a faster exploration of a larger variety of models at scale, thus significantly accelerating the project iterations. Furthermore, they will allow us to test our codebase for machine-learned plasma closures on a multi-GPU HPC environment. At the end of the allocation, we aim to publish and present our results in high-impact scientific venues in our field and make the models and codebase developed available to the research community.

DOI: https://doi.org/10.54499/2024.12865.CPCA.A1