Generalizing Shadow Mask Predictions for SPARC Plasma-Facing Components Using Machine Learning

Accurate heat-flux prediction on the SPARC tokamak Plasma Facing Components (PFCs) is critical given their complex 3-D geometries. Shadow masks—PFC regions shielded by geometry—must be included to predict surface power density and footprint. The HEAT code (Heat-flux Engineering Analysis Toolkit) has enabled precise 3-D predictions for many tokamaks by linking plasma exhaust models directly to engineering CAD [1], but individual runs take minutes to hours and are unsuitable for < 1 s applications.

To overcome this, machine learning (ML) techniques have been explored to develop surrogate models for shadow mask predictions able to run in ~1 ms. Using a feed-forward neural network (FNN) trained on a diverse database of SPARC equilibria, we successfully replicated HEAT predictions for specific PFC geometries [2], cutting computation time drastically, though the model was limited to a specific region.

To address this limitation, we now generalize the surrogate model to extend its applicability across a diverse set of PFC regions. This generalization effort involves using advanced ML architectures, particularly graph-based, which are well-suited for capturing spatial and topological relationships inherent in 3-D geometries [3]. The new graph-based method takes as input all incident field angles on the PFC mesh and outputs the shadow mask for every mesh point.

1. [1] T. Looby et al. doi: 10.1080/15361055.2021.1951532.
   
   [2] L. Antiga. Deep learning with PyTorch.
   
   [3] T. Pfaff,et al. Learning Mesh-Based Simulation with Graph Networks.