Development of a Convective–Diffusion Physics-Informed Neural Network for Thermal Analysis of Lithium Film Flow on the Surface of a Tokamak Divertor

Accurate modeling of heat transfer in nuclear engineering systems demands substantial computational resources, particularly for real-time analysis and optimization, motivating the use of artificial intelligence (AI) and deep learning (DL) as efficient alternatives. In this work, a Convective–Diffusion Physics-Informed Neural Network (CD-PINN) framework is developed to investigate steady-state heat transfer in lithium film flow along the plasma-facing surface of a tokamak divertor, where automatic differentiation is employed to evaluate residuals of the governing convection–diffusion equation and embed physical constraints directly into the training process. The model is validated using one- and two-dimensional two-layer heat conduction benchmarks, showing excellent agreement with analytical solutions, while systematic hyperparameter optimization identifies the Gaussian Error Linear Unit (GELU) activation function and the Adam optimizer as optimal for convergence and accuracy. The optimal architecture consists of 20 hidden layers with 20 neurons per layer and 5,500 collocation and boundary points, yielding steady-state temperature distributions that closely match reference solutions and demonstrate the CD-PINN’s effectiveness as a robust and computationally efficient alternative to conventional numerical solvers for complex heat transfer problems in nuclear and fusion energy systems.