Advancing IR Laser Thermography in Composites via Thermal Analysis-Informed Machine Learning

Machine learning-assisted passive thermography has emerged as a powerful tool for characterizing fatigue damage through self-heating-induced temperature hotspots, with its accuracy depending critically on the precise determination of the governing thermal properties. This work proposes a nondestructive method to simultaneously recover the transverse conductivity k_ct and the film coefficient h-two thermal parameters that are typically difficult to measure. To this end, a near-infrared laser in conjunction with a DLP-based spatial light modulator was used to create artificial temperature hotspot distributions by illuminating one side of a +/- 45 degrees glass-epoxy composite laminate. The generated surface temperature distribution on the opposite side-denoted as ground truth (GT)-was recorded by a cooled midwave infrared thermal camera. Bayesian optimization (BO) was subsequently employed to iteratively suggest the best k_ct and h values for use in a 3D FEM thermal model, which in turn generated synthetic thermal images as similar as possible to the GT. Different BO runs converged after about 10 iterations, yielding effective values of k_ct = 0.40 (W/m K) and h = 5.78 W/(m² K), from where the loss objective decreased by a factor of ca. 3 within only seven iterative rounds. A final experimental validation using these optimized parameters successfully reproduced three independent thermal images with only minor deviations, demonstrating the robustness and applicability of the proposed approach.