Terahertz Time-Domain Spectroscopy (THz-TDS) uses ultrafast lasers to emit and detect broadband, picosecond pulses with excellent signal-to-noise ratios and rapid data acquisition. Commercial spectrometers have become available and there is now great access to the technology. However, the data analysis remains complex and prone to errors due to multiple processing steps and variation in experimental setups, hindering its true breakout into industry. Machine learning, particularly the training of artificial neural networks with simulated data, has proven effective in various spectroscopic techniques, including refractive index extraction with THz-TDS. This approach allows controlled inclusion of analytical and experimental errors, enabling performant networks that are easier to characterize. We explore the use of deep neural networks for complex refractive index prediction that account for experimental and analytical errors, such as laser drift, compensating for imperfect experimental data and potentially superseding current extraction methods.
|