This paper proposes a novel frequency diverse antenna and a companion machine learning-assisted analysis of received signal spectra to determine the direction of arrival for terahertz communications across the IEEE 802.15.3d band. The antenna is an array of 3-D printed cyclic olefin copolymer (COC) 1 mm x 1 mm pillars with random heights, seven metallic scatters cut at different shapes, lengths and angles, and a WR-3 waveguide mounted off-center, all intended to significantly disrupt symmetry and generate quasi-random radiation patterns depending on the direction of arrival. Training for machine learning-assisted analysis is accomplished by Adam, an efficient optimizer applied to a convolutional neural network. Direction of arrival was determined within an average of 3.9° from ground-truth for signals from 252 to 325 GHz.