Analysis of NIF scaling using physics informed machine learning

Abstract

Over 120 DT ice layer thermonuclear (TN) ignition experiments in inertial confinement fusion (ICF) were conducted on the National Ignition Facility (NIF) in the last eight years. None of the experiments achieved ignition. In fact, the measured neutron outputs from the experiments were well below what was expected. Although experiments to fine-tune the target designs are the focus of the national ICF program, insightful analysis of the existing data is a pressing need. In highly integrated ignition experiments, it is impossible to vary only one design parameter without perturbing all the other implosion variables. Thus, to determine the nonlinear relationships between the design parameters and performance from the data, a multivariate analysis based on physics models is necessary. To this end, we apply machine learning and deep learning methods to the existing NIF experimental data to uncover the patterns and physics scaling laws in TN ignition. In this study, we focus on the scaling laws between the implosion parameters and neutron yield using different supervised learning methods. Descriptions, comparisons, and contrasts between the methods are presented. Our results show that these models are able to infer a relationship between the observed stagnation conditions and neutron yields. This exploratory study will help build new capabilities to evaluate capsule designs and provide suggestions for new designs.

The data-driven future of high-energy-density physics

“…There has been significant interest in using Bayesian inference to improve diagnostics 106 , and to synthesize observations in both focused HEDP experiments 51 and full-scale ICF experiments 50,[107][108][109] . The ultimate aims of using these methods to improve physics understanding, and the reliability of simulations in extrapolating to new designs 48 or facilities, have been addressed though machine learning 42,110 , Bayesian model calibration 50 , and transfer learning 58,111 .…”

Exploring Sensitivity of ICF Outputs to Design Parameters in Experiments Using Machine Learning

“…We define high gas fill density as any value greater than 0.6 mg/cm 3 . Expert opinion and previous work [4] indicate that, due to the significant physical differences between group I and II shots, separate analysis of each group may improve model prediction and provide insight as to the effects of the switch from high to low gas fills. In Section IV, we analyze RF performance on both groups together, while Section V includes a separate analysis of model prediction quality and feature importance rankings for each group.…”

“…Humbird et al [2] train a deep neural network (DNN) surrogate model for low-fidelity ICF simulations and apply transfer learning, a technique in which models already trained on one dataset are partially re-trained to solve different but related tasks, to obtain a surrogate model for high fidelity models and experiments. Hsu et al [4] apply ML regression methods to experimental ICF data (the same dataset analyzed in Section III) to analyze relationships between experimental outputs of interest.…”

Optimizing laser–plasma interactions for ion acceleration using particle-in-cell simulations and evolutionary algorithms

“…Due to the complexity of ultra-intense laser interactions, rather than explore the large simulation parameter space essentially by hand or some other means, instead we use an evolutionary algorithm with a series of thousands of one-dimensional (1D) particle-in-cell (PIC) simulations to optimize the laser plasma interaction. The wider field of plasma physics is beginning to embrace statistical methods for various problems such as inertial confinement fusion [19][20][21][22], magnetic fusion [23,24], x-ray production [25], laser-wakefield acceleration [26,27], and to optimize the laser focus for electron or ion acceleration experiments [28,29]. To our knowledge, the present study is the first to directly optimize laser-based ion acceleration with such an approach.…”