Extended multiphysics-informed neural network for conjugate heat transfer problems

Physics-informed neural networks have gained prominence as an innovative framework that combines machine learning and physics knowledges to solve complex physical problems. By integrating the partial differential equations of a physical system into the neural network architecture, physics-informed neural networks can approximate solutions for the system under given boundary and initial conditions. This allows the models to capture the underlying dynamics of the system while adhering to physical constraints. Despite their advantages, physics-informed neural networks exhibit performance degradation when applied to multiphysics and multi-domain problems. Specifically, conjugate heat transfer problems, which are representative of such problems, tend to exhibit low accuracy along with training instability and convergence issues when conventional physics-informed neural networks are used. Therefore, in this study, we propose a novel physics-informed neural network method called Extended Multiphysics-informed Neural Network to improve performance in solving conjugate heat transfer problems. The extended multiphysics-informed neural network framework uses distinct sub-networks designed to capture the characteristics of solid and fluid domains, enabling the accurate simulation of abrupt temperature changes at solid-fluid interfaces. Additionally, a novel training scheme incorporating physical phenomena ensures the stability and convergence of the model. The efficiency of extended multiphysics-informed neural network is validated through its application to representative two-dimensional conjugate heat transfer problems, demonstrating significant improvements in predictive performance and computational efficiency compared to existing methods. This study highlights the potential of advanced physics-informed neural network method for solving complex multiphysics problems.