The limit does exist: Evaluating digital image correlation scenarios on graph neural networks for subsurface damage detection in structures

Abstract

Digital image correlation (DIC) is a non-contact optical measurement technique used to quantify surface-level deformation. While DIC can indicate the presence of subsurface damage through surface deformation patterns, it does not provide direct insight into the extent or impact of internal damage. Machine learning models, particularly those trained on simulated data embedded with noise characteristics representative of DIC outputs, offer a potential pathway to infer subsurface damage with a quantifiable degree of confidence. However, limited research has been conducted to experimentally validate and evaluate such models using real DIC data. This study investigates the robustness of a graph neural network (GNN) trained to localize and characterize unseen or internal damage within a structural system under load. The GNN was trained on simulated data informed by representative noise inherent to real-world DIC measurements and validated against experimental specimens. A novel algorithm was developed to extract three-dimensional marker coordinates from laser scanning measurements, facilitating the alignment between DIC and finite element model (FE model) coordinate systems, a critical step in identification of damage from real-world measurement data. Following spatial alignment of the model and experimental coordinate systems, nodal results from both FE model and DIC datasets were used to map experimental measurements to FE model nodes, with smoothing applied to remaining unmatched points. The resulting data were then converted into a graph structure for damage classification. The proposed GNN-powered model was evaluated using a series of experimental specimens with varying configurations. Results demonstrate that the models were able to detect the location and severity of damage, highlighting the GNN’s ability to generalize to previously unseen noisy experimental conditions. This research provides quantified reference for practical implementation of the GNN-powered subsurface damage detection tool and shows the potential of this surrogate model being incorporated into future digital twin and structural health monitoring applications.