The Extension of Neural Networks to Compute NLTE Opacities for Multiple Elements and at Increased Fidelities

Inertial confinement fusion (ICF) is one of the primary methods that is being pursued today as a method for achieving sustained nuclear fusion for the purposes of energy generation. En route to obtaining this capability many experiments have been and will continue to be done in this pursuit; however, these experiments are costly at one million dollars or more for single experiment. Simulations thus play a key role in ICF research. These simulations are not without their drawbacks. Nuclear fusion simulations reach such temperatures and densities that normal radiation transport assumptions no longer hold and enter the regime of physics known as non-local thermodynamic equilibrium (NLTE) physics. Computations in this regime are notoriously expensive to compute. Neural networks have thus been turned to as an alternative for the standard computations of NLTE opacities. This work expands upon past work by extending a single neural network to be able to model multiple elements, and demonstrates that a median relative error for the maximum radiative temperature be at least as low as 0.566\%. Further extending the capabilities of the neural networks, it is demonstrated that neural networks can be made to predict high-fidelity spectra through the use of transfer learning achieving median relative errors of the maximum radiative temperature can be on the order of 1\%-2\% and can very likely be made better with more training data. Further, this was accomplished while achieving a computational speed up of roughly 19x for the level of fidelity used.