Geant4-Based Gamma-Radiography Detectability of Defects in Aerospace CFRP Panels: Analytical Benchmarking and Electromagnetic-Model Robustness

Yusuf Havvat

Pre-print pages 1-11

DOI: https://doi.org/10.62753/ctp.2026.05.3.3

keywords: carbon-fibre-reinforced polymer; gamma radiography; non-destructive testing; Geant4; blind detection; Monte Carlo simulation

abstract Carbon fibre reinforced polymer (CFRP) structures may contain internal delaminations, foreign inclusions, local thickness or density variations, and bond-line discontinuities that are difficult to distinguish upon external inspection. A controlled Geant4/ROOT benchmark was developed to quantify how five representative defect classes appear in single-projection gamma radiography under a common exposure budget. A 300 keV parallel gamma field rastered a 100 × 100 × 10 mm³ homogeneous-equivalent CFRP panel onto a pixelated silicon transmission plane. Seven independent four-million-history simulations comprised two matched references and five defect conditions. Defect contrast was evaluated both in a pre-specified region of interest (ROI) and by a two-sided blind scan over 141,094 position–scale candidates, with a Bonferroni global 5σ decision criterion. At four million histories, 2 mm delamination and 2 mm surface thinning produced ROI contrasts of 2.579 ± 0.147 and 2.441 ± 0.147 percentage points, with blind global significances of 16.88σ and 15.87σ, respectively. A 10% density deficit reached 7.99σ, while a 2 mm aluminium inclusion generated a sign-reversed contrast and only narrowly passed the blind threshold at 5.12σ. A 1 mm air gap inside a 1.2 mm epoxy bond line remained below the threshold (1.51σ global), despite a physically consistent positive ROI contrast. As an internal verification, the seven modelled configurations were repeated with three electromagnetic physics constructors, giving 21 independent runs with 4 × 10⁶ incident histories each. The uncollided-primary transmission agreed with the Beer–Lambert predictions within 0.36 percentage points, while the defect-contrast signs and broad hierarchy remained stable, with a maximum inter-model spread of 0.24 percentage points. The study provides a reproducible numerical ranking of idealized defect visibility and clearly separates known-location sensitivity from blind detectability; it is a simulation benchmark rather than an experimentally validated probability-of-detection qualification.