A Data-Driven Approach Towards NIF Neutron Time-Of-Flight Diagnostics Using Machine Learning and Bayesian Inference

The neutron time-of-flight (nToF) diagnostic is used to diagnose implosion dynamics during inertial confinement fusion (ICF) experiments. The primary goal of the nToF diagnostic is to extract important fusion quantities such as neutron yield, ion temperature and down-scatter ratio from neutron spectra to study the dynamics of thermonuclear fusion. In this work, we present a data-driven approach as an alternative to a physics-driven approach for nToF diagnostics at the National Ignition Facility (NIF). Instead of deriving point estimates of fusion quantities, our approach offers an approximation of the posterior distribution of the fusion quantities using a Markov chain Monte Carlo (MCMC) method. In the event of insufficient ICF experimental data, simulation outputs are needed. However, running complex simulations are computationally expensive. Hence, to speed up the data generation process, we trained a Deep Jointly Informed Neural Network (DJINN) that serves as a surrogate model to generate simulation outputs. The results of our approach and a comparison with the nToF physics-driven approach are presented in this work. Our approach is an attractive alternative to understand complex systems especially when physical models lack a complete description of the system of interest.