Using Neural Networks to Quantify Laser Filamentation for Inertial Confinement Fusion Applications

We present IrisNet, a deep Fourier Neural Operator (FNO) machine learning framework that predicts the growth of filamentation of an amplified laser pulse in a Stimulated Brillouin Scattering (SBS) gas cell, a nonlinear optical system being utilized in Xcimer Energy’s fusion approach [1]. IrisNet is trained on simulation results generated by Iris, our in-house paraxial coupled-waves code built on MFEM, while extracting physical information from the data generated by a hierarchical 1D-2D-3D parameter scan. Leveraging CUDA-accelerated PyTorch, IrisNet models the B-integral as a function of relevant physical parameters. Once trained, IrisNet is fully parameter-interpolatable, enabling instant inference of B-integral profiles for arbitrary laser input seeds, thus allowing a real-time evaluation of complex physics that would otherwise require high-dimensional PDE simulations with around one billion degrees of freedom.

[1] C. A. Thomas, M. Tabak, N. B. Alexander, C. D. Galloway, E. M. Campbell, M. P. Farrell, J. L. Kline, D. S. Montgomery, M. J. Schmitt, A. R. Christopherson, and A. Valys. 2024. Hybrid direct drive with a two-sided ultraviolet laser. Physics of Plasmas 31, 11 (November 2024). DOI: https://doi.org/10.1063/5.0221201