Transfer Learning for the Reproduction of High-Fidelity Opacity Spectra

Simulations are an important part of the inertial confinement fusion (ICF) experiment design; however, these simulations have high computational costs. One of the primary contributors to the cost is the calculation of the non-local thermal equilibrium opacities which can consume anywhere from 10% to 90%+ of the computational time. Previous studies (Deep Learning for NLTE Spectral Opacities, Gilles Kluth, 2020) demonstrate that 7x speeds up of hohlraum simulations are achievable by replacing the atomic physics calculation with a trained neural network emulator trained on several hundred thousand Cretin (Kluth) calculations. This work trained on a database of moderate fidelity DCA data for a single element.

An attractive feature of neural network emulators is that the time to evaluate the model is the same, even as the underlying training data increases in fidelity allowing for high fidelity atomic physics calculations in ICF hohlraum simulations without additional compute costs. However, generating the appropriate quantity of high-fidelity training datais extremely expensive. In this talk, we present transfer learning as a method to construct a neural network emulator for Cretin that can reproduce high-fidelity data and  provide a comparatively greater reduction in computational cost of the atomic physics calculations.