Material classification with THz radiation is typically done in transmission geometry.1 However, in many situations a reflection based classification is highly desirable. For a reflection based classification scheme, it is necessary to compensate the impact of the surface morphology on the reflection signal.2 As the surface morphology will mainly change the frequency dependent reflection pattern of the beam, we use different observation angles to improve classification based on THz reflection data. We use a THz TDS reflection setup measuring at several input and output angles. While the sample can be rotated, the transmitter can be moved on a semi-arc (see Fig. 1). We measure the reflection spectrum at different input-/output- angle configurations, which can be retrieved by an Euler transform of transmitter angle and sample angle. A measurement of a grating structure can be seen in Fig. 1. To reduce measuring time while maintaining a sufficient signal to noise ratio, we measure small angle variations around the main specular reflection. For classification we use a supervised machine learning approach based on principal component analysis for feature reduction and a support vector machine for classification.3 In this paper we present the impact of different observation angles on the classification accuracy in contrast to single-observation-angle classification, to check on the hypothesis that an increase in observation angle helps to classify a set of known materials by THz TDS reflection spectroscopy. In consequence we can estimate requirements on the observation angle and identify surface structures which will prevent classification.
|