AI and the World's Hard Problems · Nuclear fusion & energy · arXiv · 2026-10-19
ModeCast: One-Shot Prediction of Linear MHD Eigenmodes and Growth Rates in Tokamaks
Large-scale MHD instabilities in tokamaks can break plasma confinement, and evaluating a single configuration with the JOREK solver takes hours. ModeCast, a learned surrogate trained on over 45,000 JOREK simulations, predicts linear stability across toroidal mode numbers n=1 to 4, including the full spatial structure and growth rate of unstable modes. Excluding the equilibrium solve, inference on one GPU is over 10,000 times faster than multi-core JOREK linear stability runs, and it missed no unstable held-out cases. On a resistivity scan outside its training distribution, refitting only the growth rate head halved the held-out error for the internal kink growth rate.