Applying Differentiable Programming to Theoretical, Computational, and Experimental Plasma Physics

Flexible, performant implementations of Automatic Differentiation (AD) are a key enabling technology for the recent successes in machine learning (ML). AD enables accurate and fast calculations of gradients of not only neural networks, but arbitrary numerical programs. This has deep implications and creates many opportunities in wide-ranging aspects of plasma physics, from theory development and computational modeling to data analysis. In this talk, we introduce AD from a more general perspective, discuss how ML and Neural Networks fit in, and showcase applications of ML+AD towards plasma physics research. The examples include uncovering novel behavior in kinetic simulations, minimizing laser plasma instabilities for inertial fusion, learning kinetic closures to fluid equations, and performing O(>1000) parameter inverse problems for experimental data analysis. We attempt to generalize these techniques and discuss how they may be applied more broadly in plasma physics.