Robust Material Classification in Reflective Terahertz Time-Domain Spectroscopy Through Rough-Surface Training

Abstract We investigate machine-learning-based material classification in reflective terahertz time-domain spectroscopy (THz-TDS) under controlled surface morphologies. Spectra are acquired from a representative set of engineering materials with either flat surfaces or rough line profiles using a goniometric reflection setup, and a curated dataset of amplitude and unwrapped-phase spectra is prepared for training and evaluation. Classical classifiers, including support vector machines and random forests, are benchmarked across morphology splits. We find that surface morphology is a dominant factor for generalization. Models trained only on flat samples perform poorly when applied to previously unseen rough surfaces, whereas training on a combined set that includes both morphologies restores robust cross-morphology performance. On the combined data, random forests generally outperform support vector machines, and adding phase information improves results further, though careful safeguards are needed to prevent overfitting or leakage related to sample positioning. Feature-importance analysis shows a morphology-dependent shift: flat surfaces favor higher-frequency content, while rough surfaces emphasize lower-frequency bands, consistent with scattering-induced amplitude loss and reduced spectral bandwidth. These findings translate into practical guidance for reflection-based THz-TDS classification: include realistic rough surfaces during training, leverage phase information with appropriate regularization and validation, and design features that account for frequency-dependent scattering.