Cycle-GAN-based spectral reconstruction of ambient THz-TDS data for water-vapor suppression

Terahertz time-domain spectroscopy (THz-TDS) is highly sensitive to water-vapor absorption, typically requiring nitrogen purging to maintain spectral fidelity. However, this process is gas-intensive and unsuitable for high-throughput or field-deployable applications. Here, we present a cycle-consistent generative adversarial network (Cycle-GAN) that translates THz-TDS spectra between ambient-air and nitrogen-purged conditions without auxiliary humidity data. Trained on both ambient and nitrogen-purged spectra, the model achieved a frequency-averaged mean absolute error (MAE) below 0.004 for the ambient-to-purged reconstruction across the 0.2-1.8 THz band within 100 epochs, while the purged-to-ambient reconstruction exhibited higher errors up to ∼0.01 due to residual water-vapor resonances. This data-driven approach eliminates the need for gas purging and enables accurate spectral reconstruction under ambient conditions, offering a practical solution for inline quality control and compact THz sensing systems.