An imaging system that pairs terahertz time-domain spectroscopy (THz-TDS) with deep learning has been used to identify explosives and other chemicals – even when concealed under visibly opaque coverings. The researchers report pixel-level classification across eight chemicals with an average accuracy of 99.42 percent in blind tests, and 88.83 percent accuracy when detecting explosives hidden under paper, using a reflection-mode THz-TDS platform designed for stand-off measurements.
The work addresses a well-known limitation of terahertz chemical sensing: in realistic settings, spectral responses can vary with sample thickness, geometry, surface roughness, and packaging –, undermining approaches that rely on averaged frequency-domain spectra.
To overcome this, the UCLA team analyzed individual reflected terahertz time-domain pulses rather than composite spectra, using deep neural networks to classify the chemical identity of each pixel in the scanned field of view. By operating directly on pulse shapes in the time domain, the approach remains robust even when pulse timing and spacing change due to concealment or irregular sample placement.
The terahertz system operates in reflection mode and incorporates plasmonic nanoantenna arrays to enhance terahertz generation and detection. According to the authors, the setup achieves a peak dynamic range of 96 dB and a detection bandwidth of 4.5 THz within a 3-second acquisition window, using photoconductive emitter-detector pairs driven by femtosecond laser pulses. During raster scanning, a full terahertz waveform is recorded at each pixel and segmented into discrete reflection pulses.
Some of these pulses contain chemically relevant absorption information, while others originate from interfaces such as the metallic sample holder. Treating these pulses individually allows the classifier to focus on chemically stable features that persist despite variations in geometry or concealment.
The system was tested on eight chemicals: four pharmaceutical compounds – microcrystalline cellulose, dibasic calcium phosphate, mannitol, and ibuprofen – and four explosives – potassium nitrate, PETN, RDX, and TNT. In blind tests, the deep learning pipeline achieved pixel-level classification accuracies exceeding 99 percent for exposed samples and maintained high performance when explosives were concealed beneath opaque paper.
The authors position the approach as a route to stand-off, non-contact chemical imaging that combines the penetration capability of terahertz radiation with the adaptability of data-driven analysis. They also note that future developments, including faster terahertz detectors and expanded training datasets, could further improve performance under a wider range of operational conditions.
Newsletters
Receive the latest analytical science news, personalities, education, and career development – weekly to your inbox.
