Glass-fibre reinforced polymer laminates are deployed at scale across aerospace, transportation,
energy, and pressure-vessel applications, yet remain vulnerable to barely-visible impact damage
(BVID) that compromises internal load paths while leaving minimal surface evidence. Terahertz
(THz) time-of-flight imaging offers a non-contact, non-ionizing, depth-resolved probe, but its
deployment is limited by complex B-scans, expert-interpretation overhead, and the annotation cost
of supervised learning. This dissertation develops physics-informed deep-learning pipelines that turn
high-dimensional THz measurements into reliable, interpretable, and scalable inspection decisions
while reducing the manual labelling and destructive supervision required to build them. Two
synthetic corpora are produced: 1,071 BVID specimens generated by coupling Abaqus low-velocity
impact simulations to a 1D MEEP THz forward model; and 1,205 parametric Teflon-insert specimens
from MEEP alone. A transfer-learned DenseNet-121 classifier achieves 0.991 accuracy on a held-out
experimental cohort. A modality-agnostic Abaqus–MEEP-pretrained weakly supervised pipeline
converts gradient-based saliency into calibrated bounding boxes and a continuous severity output
(R2 = 0.81 on the synthetic test split), transferring without modification between THz and X-ray
micro-computed tomography data, where a YOLOv8 student detector reaches mAP@50 ≈ 0.99.
A three-tier multi-task network coupled to a calibrated physics-based interface predictor jointly
regresses defect presence, interface index, ply count, defect thickness (∼ 30 µm MAE) and diameter,
lifting depth accuracy from ∼ 14% to 91.6%, and transferring to an external experimental benchmark
with front-surface normalization alone.