Transfer learning driven design optimization for inertial confinement fusion

TL;DR

The paper addresses efficient optimization of indirect-drive ICF performance under uncertainty by integrating large-scale Hydra simulations with sparse experimental data through transfer learning. It proposes a data-driven campaign that uses a DJINN-based emulator pretrained on simulations and refined with experiments, coupled with an optimization loop guided by the expected improvement criterion $EI$. Compared to drive multiplier calibration and experimental-data-only networks, the transfer-learning approach consistently achieves near-maximum neutron yields with about 20 experiments, demonstrating faster convergence and better handling of nonlinear discrepancies between simulation and reality. The results suggest that transfer learning can enable robust, data-efficient design optimization for ICF under uncertainty, with practical relevance for guiding real experiments and broader high-energy-density physics applications.

Abstract

Transfer learning is a promising approach to creating predictive models that incorporate simulation and experimental data into a common framework. In this technique, a neural network is first trained on a large database of simulations, then partially retrained on sparse sets of experimental data to adjust predictions to be more consistent with reality. Previously, this technique has been used to create predictive models of Omega and NIF inertial confinement fusion (ICF) experiments that are more accurate than simulations alone. In this work, we conduct a transfer learning driven hypothetical ICF campaign in which the goal is to maximize experimental neutron yield via Bayesian optimization. The transfer learning model achieves yields within 5% of the maximum achievable yield in a modest-sized design space in fewer than 20 experiments. Furthermore, we demonstrate that this method is more efficient at optimizing designs than traditional model calibration techniques commonly employed in ICF design. Such an approach to ICF design could enable robust optimization of experimental performance under uncertainty.

Figures (3)

• Figure 1: The maximum observed yield, normalized by the true maximum achievable yield, as a function of the number of experiments, or shots, executed. In this study, the experimental model is defined by a diagonal matrix $B$. In red is the result of the transfer learned model, in blue is the results of calibrating cone fraction multipliers. The error bounds illustrate the variation in performance as the random matrices are varied. At the low data limit the drive calibration method achieves better performance than the TL model. However, the TL model increases in performance rapidly as data is acquired, ultimately reaching 90% of the peak yield at 20 experiments, where the calibration method learns more slowly, and reaches only around 80% of the peak yield.
• Figure 2: The maximum observed yield, normalized by the true maximum achievable yield, as a function of the number of experiments executed. In this study, the experimental model is defined by a dense random matrix $B$. In red is the result of the transfer learned model, in blue is the results of calibrating drive multipliers, and in yellow is neural network trained only on experimental data. The error bounds illustrate the variation in performance as the random matrices defining the experimental model are varied. The transfer learned model consistently learns more efficiently than the other techniques, reaching almost 90% of the true maximum yield after 15 experiments.
• Figure 3: The maximum achieved experimental yield as the number of random and optimal experiments are varied for transfer learning and cone fraction multiplier calibration techniques. Transfer learning reaches about 90% of the maximum possible yield in 20 experiments; cone fraction multiplier calibration often cannot find the maximum yield design.