Machine-learned surrogate driven exploration of the design space of a multi-shell liner target

First Light Fusion is developing magneto-inertial fusion targets for low power drivers as part of its IFE programme. Multi-shell liner targets are being designed to demonstrate neutron yield on M3, our 10 MA, 2 μs pulsed power machine.

The multi-shell concept consists of three heavy layers and two buffer gas layers, with glide planes converging on a deuterium fuel capsule. Identifying designs that are robust to magneto Rayleigh-Taylor (MRT) instabilities is a significant challenge. We simulate the fusion performance and implosion stability of these targets on M3 using our in-house magneto-hydrodynamics (MHD) code B2.

The target presents an eleven-dimensional parameter space, where each MHD simulation is computationally expensive. We explore the space using a multi-wave adaptive sampling algorithm that leverages machine-learned surrogates of the MHD simulator. The initial dataset of simulations is generated by a space-filling experimental design. Samples for subsequent waves of simulations are generated by applying Bayesian inference to determine a distribution over the design space that represents the probability of observing stable and performant liner implosions in a 2D MHD simulation. This enables sample efficient identification of the performant regions in the multi-shell liner design space for M3.