Elastodynamic imaging of voids in a PML-truncated layered solid using a deep convolutional neural network

Voids in the subsurface pose significant challenges to infrastructure stability and safety, often leading to structural failures and costly repairs. In this study, we propose a deep convolutional neural network (DCNN) framework for elastodynamic imaging of voids in a semi-infinite soil domain truncated by non-convolutional second-order complex-frequency-shifted perfectly matched layers (CFS-PML). The method employs element-wise classification to map void probabilities within the domain using elastodynamic surface responses from non-scanning type probing. The training datasets are generated using a level-set wave solver, producing input-layer features from measured surface responses and output-layer features as element grid maps indicating void probabilities. The DCNN is trained to predict the void probability in each element and reconstructs targeted voids by clustering high-probability elements. Numerical results demonstrate that, via rigorous out-of-distribution tests, the proposed DCNN can effectively detect and image voids, including those with complex shapes that were not even included in training data. The model’s performance remains stable under receiver uncertainties, including measurement noise and random tilting, with noise-trained models showing notably improved robustness. It also yields reasonable predictions under sparse receiver layouts and maintains stable performance across material-property variations. Compared with full-waveform inversion, our DCNN offers more accurate reconstructions, making void locations clearer. This study highlights the potential of integrating advanced deep-learning techniques with wave propagation models for improved subsurface exploration and characterization. • Robust deep convolutional neural network (DCNN) detects voids using elastodynamic waves. • Level-set method facilitates to model voids on a computational domain and to generate training data. • The medium is truncated by non-convolutional complex-frequency-shifted PML. • Element-wise classification can effectively predict the locations, shapes, and sizes of multiple voids. • DCNN detects deep or multiple voids more explicitly and clearly than full-waveform inversion.