Accelerating Magnetic Shadow Mask Prediction for Plasma Facing Components in Tokamaks: A Machine Learning-Based Surrogate Approach

Calculating heat flux on plasma-facing components (PFCs) in tokamaks is bottlenecked by the high computational cost of tracing magnetic field lines through complex 3D interior geometries. To address this problem, we present a machine learning surrogate that rapidly and accurately computes magnetic shadow masks. Unlike an existing approach that regresses a fixed-size output vector for a single, fixed tile geometry, we reformulate shadow mask calculation as a pointwise binary classification task over spatial coordinates and local magnetic field values, enabling a single compact model to generalize across arbitrary surface points on PFCs. Our trained neural network model achieves 97.03% IoU against the simulator-computed ground truth with a 81 times wall-clock speedup at the primary resolution.