Researchers from the University of California, Los Angeles (UCLA) have developed a chemical imaging system that combines high-performance terahertz time-domain spectroscopy with advanced deep learning techniques to accurately image, detect and classify explosives — even when those explosives are concealed or have irregular geometries.

Current methods for detecting concealed explosives and chemical threats are challenging primarily due to operational limitations. Although X-ray scanners and millimeter-wave imaging can identify suspicious shapes, they tend to lack chemical specificity.

Detection of concealed explosives using terahertz spectral imaging and deep learning. Source: Jarrahi Lab/UCLA

Meanwhile, precise chemical sensors, such as mass spectrometry or trained canine units, demand close proximity to the target, thus creating safety risks and logistical bottlenecks in crowded or volatile environments. While terahertz spectroscopy can safely see through materials like clothing and plastic, in real-world settings its chemical signatures are often distorted by packaging, surface textures and environmental scattering, thereby limiting its effectiveness.

As such, the team developed its chemical imaging system that uses plasmonic nanoantenna arrays for terahertz generation and detection, which enable the system to achieve a large dynamic range and a broad bandwidth. The team explained that unlike current approaches that rely on averaged terahertz spectra, this new system analyzes individual time-domain pulses reflected from the sample.

Specifically, these raw waveforms are processed through a custom deep learning architecture that pairs convolutional neural networks and transformers, which enable the system to disentangle the chemical signature from environmental noise and scattering artifacts.

The team validated the deep learning-enhanced imaging framework through blind testing across eight chemical species — including pharmaceutical compounds and explosives like TNT, RDX and PETN. During those tests, the system delivered a 99.42% average pixel-level classification accuracy for exposed samples, the team discovered.

Likewise, the system demonstrated generalization capabilities, maintaining average accuracy of roughly 88.83% when detecting explosives hidden under opaque paper coverings.

The team suggests that this framework could one day lead to a highly sensitive platform for security screening, pharmaceutical manufacturing and industrial quality control applications.

An article detailing the technology, “Detection and imaging of chemicals and hidden explosives using terahertz time-domain spectroscopy and deep learning,” appears in the journal Light: Science & Applications.