Neural Network Surrogates for Atomic Physics Simulations and X-ray Spectral Evaluation

Atomic physics simulations typically require significant computational resources to execute which significantly impacts the speed at which optimal fit parameters (density, temperature, plasma scale-length, etc.) can be deduced from experimental data. Here we develop a neural-network (NN) surrogate for the atomic physics simulations in order to very rapidly (10’s of ms) produce x-ray spectra based on these inputs. To further increase the speed of evaluation a principal components analysis is performed on the spectra in order to reduce the number of spectral datapoints (typically >1k in this work) to just 50 parameters. This is then used in tandem with a genetic algorithm for fitting the spectra to get approximate results which can then be checked with atomic physics simulations in the density and temperature range suggested by the NN-based genetic algorithm. We present the framework for this methodology and results from application of this technique for fitting temporally and spatially integrated x-ray spectra from proton-driven isochoric heating experiments performed on Omega EP.