Exploring Multi-fidelity Bayesian Optimization for ICF Design

Inertial confinement fusion (ICF) experiments at the National Ignition Facility are used to study high-energy density plasmas for basic science, stockpile stewardship, and fusion energy applications. Simulating these experiments requires complex coding tools, such as the LLNL-developed HYDRA [1] code, to handle laser propagation, hohlraum radiative response, capsule implosion dynamics, and to model the subsequent fusion reaction. Such codes are used to guide ICF design work; however, simulations can be costly to run, and the design space is large, spanning at least a few dozen independent parameters.

Here we demonstrate how recently developed automated tools [2] can be applied to a simplified ICF design problem. The tools leverage multi-fidelity Bayesian optimization (BO) techniques to search high-dimensional design spaces for candidate experiments. By utilizing surrogate models, the BO algorithm allows both lower and higher fidelity simulations to inform the search. We compare the search performance between neural network and Gaussian process surrogate models. We close with a discussion of applying these tools to a full ICF design problem, with the potential of neural network surrogates scaling favorably to the large ICF design space.

LLNL-ABS-851380