Towards Tearing Onset Prediction with Physics Informed Machine Learning

Improving tearing-onset prediction in tokamaks may facilitate the discovery of robust tearing-free scenarios for reactors. To this end, we have begun development of a physics-informed machine learning (ML) tearing stability metric for time-independent magnetic equilibria. A database of equilibria in cylindrical geometry is being generated and evolved linearly in M3D-C1. Inputs to the ML predictor include: observed M3D-C1 growth rates (ML training set); Δ' in the constant-psi approximation; ratios of big and small asymptotic solutions in inner resistive-MHD and outer ideal-MHD regions about the rational surface; and growth rates from asymptotic matching. Diverging stability predictions were observed between the constant-psi, zero-pressure tearing theory and full asymptotic matching in relation to a low-pressure equilibrium. M3D-C1 was able to independently match both predictions by including/excluding pressure effects. Work is progressing towards identifying stability boundaries in a multiple parameter design-space with the ML predictor. Next steps involve investigating the hypothesis that increasing classical tearing stability can improve robustness to magnetic island seeding events, and adding finite island-width effects to the model.