Implementing Physics-Informed Neural Networks for Heat Transfer Using Nvidia Modulus

Physics-Informed Neural Networks (PINNs) have emerged as a powerful tool for modelling complex physical phenomena without the need for training data. This presentation will demonstrate the use of Nvidia Modulus to construct a surrogate model based on PINNs. The aim is to compute the temperature, due to neutronic heating, within the three-dimensional wall of a pressure vessel in a nuclear fusion power plant. The focus is on a heat transfer case study that incorporates a range of thermal loads and constraints including radiation, convection, and conduction, alongside symmetry boundary conditions. The presentation highlights the capability of Nvidia Modulus to take a geometry from a computer aided design package as input and carry out uniform sampling over the geometry's surface and volume. It also describes the use of custom sample points to apply the necessary heat sources. Additionally, the authors outline the process of formulating the governing partial differential equation for heat conduction, as well as establishing the thermal-related boundary constraints. A significant emphasis is placed on the methodology for tuning the weights of each loss term, which is critical for optimizing the performance and accuracy of the surrogate model. Validation using standard finite element simulation demonstrates alignment between the results of finite element simulations and the developed PINN. The presentation will be of interest to engineers and scientists that are considering the use of PINNs for their work.