Meta-Learned Bayesian Optimization for Energy Yield in Inertial Confinement Fusion (Papers Track)
Vineet Gundecha (Hewlett Packard Enterpise); Ricardo Luna Gutierrez (Hewlett Packard Enterprise); Sahand Ghorbanpour (Hewlett Packard Enterprise); Rahman Ejaz (University of Rochester); Varchas Gopalaswamy (University of Rochester); Riccardo Betti (University of Rochester); Avisek Naug (Hewlett Packard Enterprise); Paolo Faraboschi (Hewlett Packard Enterprise); Soumyendu Sarkar (Hewlett Packard Enterprise)
Abstract
With the growing demand for clean energy, fusion presents a promising path to sustainable power generation. Inertial confinement fusion (ICF) experiments trigger nuclear reactions by firing lasers at a fuel target, typically composed of deuterium and tritium. These experiments are costly and require complex optimization of the laser pulse shape across multiple shots to maximize energy yield. Even though Bayesian Optimization (BO) has been commonly used to optimize such expensive scientific experiments, vanilla BO methods do not leverage prior knowledge of the function from simulations or past experiments and fail to achieve high sample efficiency. In this work, we adapted and explored BO meta-learning techniques for ICF that either meta-learn the BO surrogate model, the acquisition function, or both from simulations. Our results demonstrate that the three meta-learning techniques we investigated, Meta-Learning Acquisition Functions for BO (MetaBO), Rank-Weighted Gaussian Process Ensemble (RGPE), and Neural Acquisition Processes (NAP), drastically reduce the number of experiments needed to achieve a satisfactory yield in ICF simulations.