EM-RBN: A Physics-informed Neural Network Algorithm Based on Integrated Networks and Dynamic Weight Generation

Because of the low convergence efficiency and prediction accuracy of clasical physics-informed neural network caused by spectral bias, a physics-informed neural network algorithm (EM–RBN) based on integrated networks and dynamic weight generation is proposed. This algorithm captures cross-frequency features by designing a multi-layer perceptron based on data augmentation transformation and a radial basis sub network. Additionally, a dynamic weight generation module is introduced to learn and adjust the output weights of sub-networks in real time, thereby balancing the contributions of high and low frequency features. Experimental results show that the EM–RBN method achieves a maximum prediction accuracy improvement of two orders of magnitude compared with existing physics-informed neural network algorithms across different numerical problems. The code of this paper is open-sourced at https://github.com/jia-chen25/EM-RBN.