Enhanced Physics-Informed Neural Networks with Residual Architecture for Accurate Heat Conduction Predictions

In this paper, a Physics-Informed Neural Network (PINN)-based model, enhanced with a Residual Network(ResNet) architecture, is proposed and validated for solving the two-dimensional heat conduction equation. By embedding the physical constraints of the heat conduction equation directly into the training process and leveraging the ResNet’s capability to model complex functions efficiently, the model achieves accurate temperature field predictions under unsupervised learning conditions. The proposed model demonstrates a prediction error of less than 1%, showcasing its high accuracy. Additionally, the model exhibits strong generalization capability at different times, effectively predicting temperature distributions even in scenarios not explicitly represented during training.