U-Net-based deep learning enabling denoised spectroscopy for terahertz quantum cascade laser

Terahertz quantum cascade lasers (THz QCL) based on self-mixing offer great potential for high-resolution spectroscopy and imaging, yet their practical application is hindered by the low accuracy and signal-to-noise ratio (SNR) using traditional Fourier techniques. To address this issue, we propose a hybrid denoising framework combining wavelet pre-processing and a peak-aware U-Net network to restore multimode THz QCL spectra with enhanced SNR and reduced error factor. The wavelet transform first removes high-frequency noise, while the subsequent U-Net reconstructs spectral details and enhances peaks via a dedicated peak-aware mechanism. Experimental results demonstrate that our method achieves an SNR enhancement by ∼ 10 dB, a determination coefficient of 0.979, and a reconstruction error of 3.7 × 10^-5, significantly outperforming the original data. Validation on methanol gas transmission spectra further confirms the ability of the model to accurately recover the characteristic spectral information. This work illustrates the strong capability of deep learning-assisted signal processing in advancing THz spectroscopic sensing toward high accuracy and robustness.