In the research of inertial confinement fusion (ICF), the computational complexity of time-dependent non-local thermodynamic equilibrium (NLTE) models significantly limits simulation efficiency of large-scale evolution simulations. This study proposes a neural network surrogate model to accelerate NLTE spectral property calculations in gold plasmas.

To accurately capture the time-dependent characteristics of the plasma, we propose a neural network surrogate model to accelerate NLTE spectral calculations in gold plasmas. The model captures time-dependent dynamics by incorporating the time step, alongside current ionization and orbital populations, into its input vector. Architecturally, we employ a multi-branch parallel residual network (ResNet) to process multi-scale physical quantities simultaneously. The network features three dedicated branches for predicting updated ionization populations, orbital populations, and absorption/emission coefficients. To guarantee physical accuracy, we apply tailored data preprocessing (e.g., logarithmic scaling and threshold cutoffs), integrate physics-informed constraints into the loss function, and enforce strict constraint mechanisms during post-processing. By significantly reducing computational time while preserving physical fidelity, this surrogate model offers a highly efficient alternative to traditional time-dependent NLTE modules in future ICF radiation hydrodynamic simulations.
 

ICF,Machine Learning,NLTE