Synthetic-data-trained deep learning enables quantitative terahertz metrology in pharmaceutical coatings

Abstract

Terahertz time-domain imaging provides non-destructive access to buried layers in coated materials, yet quantitative inversion of reflection-mode measurements remains challenging because of ill-posed nature of waveform interpretation and the scarcity of experimentally labelled training data. Here we report a physics-informed deep learning framework trained exclusively on synthetic terahertz waveforms generated from an electromagnetic multilayer model. By incorporating domain randomisation to emulate realistic experimental variability together with a physics-guided parameterisation of coating properties, the model transfers directly from simulation to experimental reflection-mode measurements without labelled data or post hoc calibration. Using pharmaceutical film coatings as a representative system, we quantitatively recover optical thickness and refractive-index from measured terahertz waveforms, with synthetic validation demonstrating accurate recovery over the complete parameter range, including thin-coating regimes where conventional peak-finding fails because temporal peak separation is lost. The framework generalises across independently manufactured batches and measurements acquired on different days. These results demonstrate a scalable route towards label-free quantitative terahertz metrology and highlight the potential of synthetic-data-trained, physics-informed learning for experimentally constrained inverse problems.