AI and the World's Hard Problems · Nuclear fusion & energy · arXiv · 2026-10-20
Reinforcement Learning Controllers for Fusion Scenario Development
Designing tokamak discharge scenarios is iterative and labour-intensive, since each candidate needs controller tuning and simulation. The authors train one goal-conditioned reinforcement-learning controller in advance, specified only by target waveforms, using a free-boundary equilibrium and transport simulator. On an EAST plasma-current ramp-up, the controller formed a divertor in 86% of 1024 test scenarios and reached the end of the episode in a diverted state in 58%, with failures concentrated at high current density.