AI-Driven Void Detection Using SPECFEM3D Wave Propagation Simulations

This study introduces a deep learning framework for detecting underground cavities using synthetic seismic displacement data generated by SPECFEM3D. Traditional inversion methods are computationally intensive and often limited in resolution, motivating the use of a 3D convolutional neural network (3DCNN) to learn spatial-temporal features from sparse surface sensor data. The model was trained to predict a continuous 3D probability matrix indicating cavity presence, which was later post-processed for binary classification and evaluation. High accuracy was achieved across regression and classification metrics. Comparing simulated wave responses from predicted cavities to the original sensor data, confirmed physical consistency. Additional blind tests demonstrated robustness under varying conditions. Overall, this approach offers a promising, efficient alternative for subsurface anomaly detection with strong predictive performance and physical reliability.