Material classification

Terahertz Time-Domain spectroscopy (TDS) allows for automated classification of material samples by training a a machine learning-based classifier. Our work is particularly concerned with classification on the basis of reflective TDS measurements. The investigated samples may have a surface roughness that interferes with the relatively short-wave terahertz radiation (see also Influence of surface structures).

If a classifier is trained exclusively with reflective measurements of smooth-surface samples, this perturbation is not accounted for. The classification performance will be significantly reduced, potentially up to complete failure, if such a classifier is only tested but not trained with rough surfaces. We could show that in turn combining the data of samples with different surface topologies, i.e. rough and flat specimen, for training the algorithm significantly improves the reliability of the material classification.