Detection and imaging of chemicals and hidden explosives using terahertz time-domain spectroscopy and deep learning

Detecting concealed explosives and chemical threats constitutes a critical challenge in global security, yet current technologies often face significant operational limitations. While X-ray scanners and millimeter-wave imaging can efficiently identify suspicious shapes, they frequently lack the chemical specificity. Conversely, precise chemical sensors—such as mass spectrometry or trained canine units—require close proximity to the target, creating safety risks and logistical bottlenecks in crowded or volatile environments. Terahertz spectroscopy has long been viewed as a promising solution to this dilemma, offering the ability to penetrate opaque materials like clothing, paper, and plastic without causing ionization damage. However, conventional terahertz methods often struggle in real-world conditions, where the spectral fingerprints of chemicals are easily distorted by packaging materials, surface roughness, and environmental scattering.

 

In a new article published in Light: Science & Applications, a team of researchers, led by Professors Mona Jarrahi and Aydogan Ozcan from the University of California, Los Angeles (UCLA), USA, has addressed these challenges by developing a robust chemical imaging system that integrates high-performance terahertz time-domain spectroscopy with advanced deep learning techniques. This system is designed to accurately image, detect and classify explosives even when they are concealed or possess irregular geometries.

 

A key enabler of this technology is the use of plasmonic nanoantenna arrays for terahertz generation and detection, which allow the system to achieve a large dynamic range and broad bandwidth. Unlike traditional approaches that rely on averaged terahertz spectra, this system analyzes individual time-domain pulses reflected from the sample. By processing these raw waveforms through a custom deep learning architecture combining convolutional neural networks and transformers, the system can disentangle the unique chemical signature from environmental noise and scattering artifacts.

 

The efficacy of this deep learning-enhanced imaging framework was validated through rigorous blind testing across eight distinct chemical species, including pharmaceutical compounds and high-priority explosives such as TNT, RDX, and PETN. The system achieved a remarkable average classification accuracy of 99.42% at the pixel level for exposed samples. Crucially, it demonstrated robust generalization capabilities in challenging scenarios, maintaining an average accuracy of 88.83% when detecting explosives concealed under opaque paper coverings—a task where conventional spectral methods typically fail. This framework offers a highly sensitive platform for rapid, stand-off chemical imaging, with transformative potential for security screening, pharmaceutical manufacturing, and industrial quality control.

 

The authors of this work are Xinghe Jiang, Yuhang Li, Yuzhu Li, Che-Yung Shen, Aydogan Ozcan, and Mona Jarrahi of UCLA Samueli School of Engineering.

---

---