An enhanced hybrid adaptive physics-informed neural network for forward and inverse PDE problems
摘要
Physics-informed neural networks (PINNs) have emerged as a powerful tool for solving partial differential equations (PDEs) in various scientific and engineering applications. PINNs integrate the PDEs into the loss functions of neural networks through automatic differentiation, which comprises a weighted combination of the PDE residuals, boundary and initial constraints, and observed data. However, the accuracy and efficiency of PINNs need substantial improvement to meet the requirements across a wider range of challenging problems. In this paper, we propose a hybrid adaptive (HA) sampling method and a feature embedding layer for PINN. The spatiotemporal points for PINN training (called residual points) are resampled during iteration procedure via the proposed HA method. The HA method ensures randomness in the selection of points while also prioritizing those with large PDE residuals, thereby ensuring efficient training by focusing the network’s learning on areas where the model’s predictions are less accurate. Moreover, the vanilla PINN architecture is further enhanced by a feature embedding layer, which transforms the raw input into a higher-dimensional space, enabling the network to better capture complex relationships underlying in PDEs and improve its fitting ability. The numerical experiments on a variety of forward and inverse PDE problems have shown that the proposed methods significantly improve the accuracy and efficiency of PINN with reduced reliance on the number of sampling points. The L2 relative error of vanilla PINN can be reduced by approximately 1