Abstract

Nuclear fusion holds the promise of limitless clean energy and would solve many of the world's grand challenges. The most promising approach to date uses tokamaks, but we have not yet been able to sustain plasmas at the temperatures, pressures, and durations needed to make fusion power viable. This is due to the stochastic, nonlinear, unstable nature of plasmas; the expense of running experiments on the real device; and the poor fidelity of simulators.
Reinforcement learning, Bayesian optimization, and other AI approaches have become increasingly capable, which makes them an appealing option for nuclear fusion. Unfortunately, RL has been less successful in stochastic and partially observed problems and both RL and BO struggle when given only a few experiments. In this talk I will present several algorithmic innovations to address these issues. I will address achieving better calibration when doing uncertainty quantification with neural network ensembles; present a method for using priors from learned dynamic models to make BO possible with few experiments; and describe our RL pipeline incorporating the dynamics models and UQ. Finally, I'll show results of our recent experiments on tokamaks.