A multi-domain physics-informed neural network for transient thermal analysis of a Tokamak divertor

Abstract

In this study, a Multi-Domain Physics-Informed Neural Network (MD-PINN) framework is implemented to model a multi-layer transient heat conduction problem with complex boundary conditions in an upgraded mockup of the Pakistan Spherical Tokamak divertor configuration. Automatic differentiation is employed to compute the spatial and temporal derivatives of the temperature field, which are essential for enforcing the residuals of the governing equations, initial conditions, boundary conditions, and interface conditions in the loss function. This approach rigorously embeds the physical constraints of the problem into the neural network’s training process. The model is validated against analytical solutions for one-dimensional and two-dimensional heat conduction problems, demonstrating strong agreement. It is then applied to predict transient temperature distributions in a two-dimensional graphite divertor mockup, with results closely matching the reference solutions. While the approach effectively captures the underlying physics, the hyperparameter tuning and training phase requires significant computational time. However, once trained, the model enables near-instantaneous predictions, making it well-suited for real-time feedback applications in plasma-facing components. This study paves the way for future extensions toward an autonomous, artificial intelligence-driven simulation tool for complex, coupled multi-physics problems involving three-dimensional geometries and nonlinear materials.