Neural network surrogate models for absorptivity and emissivity spectra of multiple elements

Abstract

The emissivity and absorptivity of radiation by matter are fundamental physical parameters required to adequately simulate inertial confinement fusion experiments. We present novel neural network models for predicting these interaction coefficients. This research extends the methods used by Kluth et al. (2020) that made similar predictions for a single chemical element, krypton. Our models are based on fully-connected or convolutional autoencoders coupled with a deep jointly-informed neural network (DJINN) to predict the emissivity and absorptivity of a given element and temperature/radiative field. We show that the previous work could not be directly extended to lower atomic number elements due to a thresholding effect on small values caused by a log10 transformation. Thus, in order to create a multi-element model, or even a single-element model, with low atomic number elements a different transformation is necessary. Utilization of a cube-root transform enables the creation of a multi-element model can achieve mean relative errors between 1% and 2%. Our work demonstrates that a single neural network can predict the results of atomic physics calculations, but additional work is necessary to consider mixtures of elements or a wider range of elements than those used in our study.