GPU-accelerated algorithm for measuring terahertz spectral characteristics

Terahertz-based non-destructive testing utilizes a point-scanning imaging process to capture comprehensive regional information through high-density grid sampling. Accurate estimation of characteristic parameters such as refractive index, dielectric constant, and absorption coefficient requires addressing a multi-parameter numerical problem with micrometer-scale precision in sample thickness. Given the demands of high frame rate imaging, purely CPU-based computing frameworks are insufficient for the required rapid processing. Consequently, this paper introduces a GPUaccelerated terahertz spectroscopic characteristic parameter measurement algorithm using CUDA technology. The system enhances performance by streamlining data transfers between devices and performing batch Fourier transformations on multiple short one-dimensional sequences. It also leverages shared memory to efficiently execute distributed detection of signal peaks and spectral matching. Furthermore, a fitness function based on the distribution of local extrema across multiple consecutive data buckets is proposed, facilitating rapid comparison and optimization of results. Experimentally, the system was tested on both simulated data of single and double-layer media and real data from the MenloSystems terahertz time-domain spectrometer. Experimental results indicate that NVIDIA GeForce 30 series GPU acceleration increases processing speed by 5 to 8 times relative to traditional CPU methods. Additionally, by exploring and sampling in a more refined parameter space while maintaining a computation speed of 100ms per measurement point, the accuracy of refractive index determination improved by 80%. These achievements not only underscore the potential of GPU acceleration in enhancing terahertz spectroscopy performance but also provide substantial support for the continued advancement of related technologies.