Millimetre-Wave Non-Destructive Evaluation of Concealed Interfaces in Multilayer Assemblies

Large-scale assemblies in aerospace and manufacturing often contain concealed interfacial features, including unintended air gaps, non-contact regions, thickness non-uniformities, and imperfect interface closure, that can degrade structural performance and reliability. Because such features are difficult to assess after assembly, there remains a need for an inspection method that can operate from one side, without contact, and provide quantitative information about buried dielectric interfaces. The aim of this research is to develop a non-contact millimetre-wave inspection framework for quantitative estimation and mapping of thickness and air gap variation in multilayer planar assemblies.
This thesis makes three contributions. First, it develops a free-space dielectric characterisation method for W-band measurements based on joint inversion of complex permittivity and setup nuisance parameters. By combining air normalisation with a slab model, the method allows dielectric retrieval under reference plane uncertainty, stand-off variation, and coupling mismatch. Second it establishes a one-sided reflection methodology for conductor-backed and multilayer dielectric structures, showing both where broadband S11 inversion can recover thickness reliably, and where recovery becomes unstable because of stand-off variation, imperfect interface closure, non-parallel geometry, and coupling between thickness and air gap.
Third, it develops a physics-informed data-driven framework for multilayer thickness and air gap prediction by combining analytical data, full-wave CST simulation, and physical measurements in a dual-stream model trained on spectral and time-domain information.
The results show that broadband W-band measurements contain sufficient information for quantitative inspection of concealed dielectric interfaces, but direct analytical inversion become increasingly fragile as nuisance effects and parameter coupling intensify.
The thesis contributes to knowledge by demonstrating practical dielectric retrieval in a realistic free-space fixture, defining the capability and limits of one-sided
reflection inversion for multilayer assemblies, and showing that progressive multi-domain training improves quantitative prediction on physical measurements without requiring a very large labelled experimental dataset.