ACHIEVING TRANSFERABLE MULTIVARIATE SPECTRAL CALIBRATION MODELS - DEMONSTRATION WITH INFRARED-SPECTRA OF THIN-FILM DIELECTRICS ON SILICON

Multivariate calibration methods are very useful in improving the precision, accuracy, and reliability of quantitative spectral analyses. In this study, the transfer of partial least-squares (PLS) calibration models between spectrometers was investigated for the quantitative infrared analysis of borophosphosilicate glass (BPSG) thin films on silicon wafers. In the determination of phosphorus, boron, and thickness for BPSG films, sensitivity studies showed that detector nonlinearity, frequency accuracy, incident angle, and variations in purge were important parameters to control to obtain transferable calibrations. The difficulty in achieving a transferable calibration model was found to increase as the complexity of the calibration model increased. A combination of controlling sensitive experimental parameters, selecting appropriate frequencies, and subtracting the spectra of purge gases improved the standard error of prediction for phosphorus determination from 0.75 to 0.18 wt % when predicting on spectra collected on a spectrometer different from that upon which the calibration model was based. This improvement in prediction ability resulted in a PLS calibration model that was transferable between two spectrometers. It has been shown that sensitive outlier detection methods can be used to identify model transfer problems. The ability to achieve calibration models that are robust to spectrometer variations implies that these calibrations can also be robust to spectrometer drift.