On-chip intelligent terahertz sensing platform for real-time intelligent biochemical identification

Abstract

This study presents a groundbreaking on-chip intelligent terahertz (THz) sensing platform for real-time biochemical identification, integrating THz time-domain spectroscopy (THz-TDS) with a low-power, embedded deep learning system based on a field-programmable gate array (FPGA) chip. Utilizing a hardware-software co-design approach, the platform deploys a custom-designed deep convolutional neural network model (DCNNM) onto an FPGA chip. This integration enables the real-time, high-specificity identification of 12 distinct biochemical compounds, including amino acids, proteins, and saccharides, achieving a remarkable accuracy of 96.67 % with a sub-second latency of 320 ms per spectrum. Critically, the platform overcomes the challenge of identifying spectrally ambiguous substances lacking distinct THz peaks (e.g., proteins and saccharides) through the proposed DCNNM's advanced feature extraction capabilities, demonstrating transformative potential for point-of-care diagnostics and drug screening. The optimized DCNNM architecture, featuring only 14.7 million parameters, ensures efficient FPGA implementation. The entire system operates at an ultra-low power consumption of 2.97 W, eliminating reliance on offline processing and enabling truly portable, embedded analysis. This innovative approach offers a powerful solution for medical diagnostics, drug screening, and biochemical research, providing a rapid, intelligent, and highly accurate method for biochemical substance identification.