Probabilistic delamination estimation in composite laminates from thermography using a deep-learning surrogate
Year: 2026
Type: article
Abstract: Quantitative sizing of delamination damage in composite laminates from active infrared thermography (IRT) is challenging due to diffusion-driven blurring, depth attenuation, and non-uniform heating. We propose a surrogate-assisted, likelihood-free Bayesian inversion framework that combines an experimentally calibrated finite-element (FE) heat-transfer model, a spatio-temporal deep-learning surrogate that maps a 3D defect mask to a surface temperature evolution, and Approximate Bayesian Computation with Subset Simulation (ABC–SubSim) to infer posterior distributions of delamination location and size. The framework is evaluated on held-out FE simulations and on experimental thermograms collected under three inspection configurations: normal transmission (NT), oblique transmission (OBT), and oblique reflection (OBR). On held-out FE cases, ABC–SubSim concentrates within seven subset levels and recovers in-plane and through-thickness bounds with more than 90% precision and less than 10% posterior uncertainty. On experimental data, NT provides the best overall estimates of damage location, with in-plane precision of 80.24%–99.80% and through-thickness precision of 71.87%–94.00%. OBT yields intermediate performance, while OBR exhibits the largest in-plane errors. The surrogate enables rapid posterior exploration and the results show that inference quality depends strongly on the heating configuration and the accuracy of the FE model used to train the deep-learning surrogate.
Source: Composite Structures
Institution: University of Illinois Urbana-Champaign
Language: English
Subfield: Mechanics of Materials
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $4,660
Award: G1554672200
OpenAlex ID: W7203820342
