A Hybrid PSM-CNN-LSTM Network for Scattered Acoustic Field Prediction

Accurately predicting scattered acoustic fields is crucial in nondestructive testing. The analytical methods achieve high computational accuracy, but are only applicable to problems with regular geometries and well-defined boundary conditions. While numerical methods can handle complex structures, they incur significantly higher computational costs. Neural networks offer computational efficiency but often fail to maintain accuracy in complex problems due to their unconstrained physical representation. This work aims to combine physics-based models and deep learning to improve accuracy and computational efficiency. This work proposes a hybrid PSM-CNN-LSTM (PCL) network for solving the scattered acoustic field with complex defects. PSM computes spatial derivatives using Fourier transforms, offering spectral convergence and capturing the frequency-domain behavior of wave equations. CNN extracts spatial features, while LSTM models temporal dynamics. This model fuses frequency features based on PSM with raw data to jointly capture spatial derivatives and temporal signal features. We employed the PCL model to predict the scattering wavefields induced by irregularly shaped defects and benchmarked its performance against PSM and CNN-LSTM. The accuracy of the proposed model was substantiated through numerical simulations, while its reliability was further corroborated by experimental validation, jointly demonstrating the effectiveness of the approach.