Multivariate calibration, as a mechanism for modeling the relationship between a substance’s spectroscopic profile and analyte concentration, allows for real-time analysis through spectroscopic measurement techniques such as near infrared (NIR) spectroscopy. However, since spectroscopic profiles also depend on measurement conditions, such as humidity, instrument drift, manufacturer, etc., model performance degrades when original (primary) conditions are moderately different than novel (secondary) measurement conditions. Model updating allows for domain adaption between primary and secondary domains by tuning the multivariate calibration model to the differences of the domains. In this presentation, we analyze the traditional method of model updating, where new samples with recorded analyte value (labeled reference values) are required to represent the secondary domain, is compared to two new processes. One needs no secondary labels (transductive), and the other includes a few secondary labels with the unlabeled secondary samples. In both of these new processes, the updated models are tuned specifically for the samples to be predicted, thereby assuring domain matching. Furthermore, all model updating situations require model selection based on multiple metaparameters and the approaches presented are based on two or three tunable metaparameters for model selection. The model diversity and prediction similarity (MDPS) model selection framework is applied, showing consistent and robust selection of models achieving at or below the first quartile of all generated models. This presentation shows that these model updating and selection frameworks rival the performance of a complete model recalibration, without the additional expense of novel sample collection from the secondary domain as well as obtaining the time consuming sample analyte reference values. Results are presented for four NIR spectral data sets.<br/>

Model updating and selection frameworks without reference values (calibration standards)

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.