Improving the transfer ability of calibration model for terahertz spectroscopy

Inter-instrument variability and measurement condition changes pose great challenges for the maintenance of terahertz calibration models. The predictive accuracy of the trained calibration models is questionable for newly measured data sets. Obviously, this problem may hinder practical applications of terahertz spectroscopy. To tackle this problem, for the first time we explore to improve the transfer ability of calibration model for terahertz spectroscopy via spectral space transformation. This method tries to minimize the spectral inconsistency induced by the changes of instruments or measurement conditions via spectral transformation between two spectral spaces spanned by the spectra of pure chemical transfer samples. Furthermore, multi-way partial least squares methods were carried out and compared with this approach based on a terahertz spectra dataset. Experiments show that spectral space transformation outperforms other methods in the sense of prediction accuracy and applicability; with this method the performance of the master model applied to slave dataset is consistently enhanced, which proves their efficacy and usability in real-life applications.