Data-driven Techniques for Time Domain Decomposition of Plasma Physics Simulation

We present the analysis of data-driven techniques for dimensionality reduction of plasma physics simulation. Surrogate models derived from limited output of such codes offer an approach to lower the computational burden of prohibitive calculations in high spatial or temporal regimes. We target the application of these models towards representation of time domain features in plasma edge simulations of the tokamak scrape-off layer boundary obtained from SOLPS-ITER. Starting from low-rank matrix approximation, we showcase a principled parameter tuning for selection of sampled simulation snapshots, construction of operators for time advance, and estimation of reduced model accuracy. We consider two use cases for the proposed techniques: extraction of best-fit linear dynamics from nonlinear data and identification of CUR decompositions as an efficient alternative to the SVD. In the former case, dynamic mode decomposition provides a method to segregate timescales and extrapolate forward stable features. In the latter case, skeleton decomposition provides an interpretable representation of matrix data optimized over physical coordinates and recorded timesteps. We discuss the implementation of these techniques for the acceleration and compression of plasma physics simulation.