Optimizing a HYDRA Test Problem: Lessons for Bayesian ICF Design Optimization
Shailaja Humane, University of Michigan
Inertial confinement fusion (ICF) experiments rely on complex multi-physics simulations such as the LLNL-developed HYDRA [1] to guide design work. However, these simulations can be expensive and have up to a few dozen design parameters, making the search for an optimal design difficult. Recently developed automated tools [2] utilize multi-fidelity Bayesian optimization to search these high-dimensional design spaces for candidate experiments. Using surrogate models, the Bayesian optimization algorithm allows both lower and higher fidelity simulations to inform the search.
These tools require tuning to efficiently search for high yield ICF designs. In an 8-parameter test problem, we investigate how the choice of cost function (specifying the relative costs of different fidelities) and acquisition function affect the multi-fidelity optimization performance. We benchmark the algorithm’s performance against uniform random sampling of the parameter space to identify the conditions under which the algorithm provides a benefit to optimization. We demonstrate that level of fidelity correlation can significantly influence the tuning needed to improve efficiency. We discuss the relevance of our findings to future HYDRA-based automated searches.
LLNL-ABS-863994
[1] Marinak, M. M., et al., Phys. Plasmas 8, 2275 (2001)
[2] Thiagarajan, J. J., et al. ICML (2022)