A More General Electromagnetic Inverse Scattering Method Based on Physics-Informed Neural Network

Based on the computational framework of physics-informed neural networks (PINNs), an unsupervised deep learning method is developed for inverse problems, which features good accuracy, high efficiency, and good generality. When considering the case of multifrequency inversion, a frequency scale factor is introduced to address the scale difference problem brought by the different frequency terms and the occurrence of gradient explosion during the network training. In addition, to improve the efficiency and accuracy, a dynamic sampling strategy is proposed. Four numerical examples and one experimental example are considered to validate the effectiveness of the proposed method. The inversion results show that the proposed PINN method achieves good accuracy, efficiency, and generality, especially for electrically large and high-contrast scatterers. Moreover, the method shows good robustness against noise. Compared with traditional data-driven deep learning methods, the proposed method is efficient because it operates in an unsupervised manner and exhibits good generalization across different inversion tasks. Compared with traditional quantitative inverse scattering algorithms, the proposed method can overcome their limitations in dealing with extremely high-contrast or electrically large targets. In general, the proposed PINN not only inherits high inversion quality when compared with traditional deep learning methods but also has better generality than the traditional inverse scattering methods.