Computational framework for guided wave NDT to identify delamination of arbitrary geometries in 3D layered...

We propose a convolutional neural network (CNN) framework designed to identify subsurface delamination within a 3D anisotropic composite structure using elastodynamic wave measurements. The network is optimized through a grid-wise classification approach, estimating the probability of internal defects at each grid element and enabling precise spatial localization of damage. To construct the training dataset, we perform high-fidelity simulations in ANSYS Mechanical, modeling wave propagation through orthotropic media, layered between isotropic materials, with randomly generated delaminations. Each training sample consists of input-layer features (i.e., time-domain wave responses recorded by sensors like laser Doppler vibrometer (LDV)) and output-layer features (i.e., damage probability maps on the fixed background grid at the interface between layers). The background grid is unaltered during the data generation process. The proposed CNN interprets sensor waveforms to detect and localize internal delaminations based on predicted damage probability maps. An alternative evaluation metric-Class-Weighted Accuracy (CWA)-is proposed for assessing the performance of the CNN to resolve a highly-imbalanced binary classification problem. The framework is validated using in-distribution test datasets, as well as out-of-distribution damage cases, which are unseen during training, demonstrating high accuracy in detecting delaminations of varying shapes, quantities, and locations. Numerical experiments show that a greater number of sensors and larger training datasets improve the detection accuracy of the CNN; training the CNN with noisy data results in more reliable performance under noisy measurement conditions than training without noisy data; high accuracy of the CNN is achieved even when only out-of-plane measurements are used; using a finer grid resolution effectively improves defect localization accuracy and boundary reconstruction; and high performance stability is maintained under material and geometric uncertainties; even nontrivial-shaped delaminations can be identified without prior knowledge of their shape or quantity. This approach enables robust and scalable nondestructive evaluation in composite materials without requiring prior knowledge of internal defect characteristics.