# Transfer Learning as a Method to Reproduce High-Fidelity NLTE Opacities in Simulations

**Authors:** Michael D. Vander Wal, Ryan G. McClarren, Kelli D. Humbird  
**arXiv:** [2205.14520](https://arxiv.org/abs/2205.14520)  
**Date:** 2022-05-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.14520v1/

## Key points

- Simulating high-energy density physics requires non-local thermodynamic equilibrium (NLTE) opacity data
- NLTE opacity data can be expensive to produce at high-fidelity
- Researchers have explored using neural networks to replace standard calculations of low-fidelity data and train them to reproduce artificial high-fidelity opacity spectra
- A novel neural network architecture was trained through transfer learning to reproduce high-fidelity krypton spectra for use in simulations
- The neural network achieved a relative percent error of the peak radiative temperature of the hohlraum between 1% and 4%, while also achieving a 19.4x speedup
- Physics informed neural networks (PINNs) may provide higher quality results but would require a custom training loop run concurrently with radiation-hydrodynamics software packages like Cretin
- There could be further improvements in the architecture by introducing convolutional networks or AI-based methods for determining which network should take high-temperature inputs versus low-temperature inputs
- Simulating NLTE opacities remains critical for understanding many integrated hohlraum simulations on some of the world's most powerful supercomputers
- Calculation of spectral absorptivity and emissivity alone can consume up to ninety percent of computation time by way of simulation

## Summary

Simulating high-energy density physics often requires non-local thermodynamic equilibrium (NLTE) opacity data, which can be expensive to produce at high-fidelity. To address this issue, researchers have explored using neural networks to replace standard calculations of low-fidelity data and train them to reproduce artificial high-fidelity opacity spectra. In a recent study, a novel neural network architecture was trained through transfer learning to reproduce high-fidelity krypton spectra for use in simulations. The results showed that the neural network achieved a relative percent error of the peak radiative temperature of the hohlraum between 1% and 4%, while also achieving a 19.4x speedup. However, there is still room for improvement in the accuracy and efficiency of these models. For instance, physics informed neural networks (PINNs) may provide higher quality results but would require a custom training loop run concurrently with radiation-hydrodynamics software packages like Cretin. Additionally, there could be further improvements in the architecture by introducing convolutional networks or AI-based methods for determining which network should take high-temperature inputs versus low-temperature inputs. Despite these challenges, simulating NLTE opacities remains critical for understanding many integrated hohlraum simulations on some of the world's most powerful supercomputers. The calculation of spectral absorptivity and emissivity alone can consume up to ninety percent of computation time by way of simulation. This work was performed under the auspices of the U.S Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 and released as LLNL-JRNL-834755-DRAFT.