A spectral calibration forms a prediction model by accurately representing the relationship between measured sample spectra and analyte concentration. Impeding real-time analysis is that spectral profiles depend on measurement conditions, such as humidity, temperature, instrument drift, manufacturer, etc. as well as sample composition relative to analyte and other species amounts and the physicochemical sample matrix effects from inter- and intra-molecular interactions.  Thus, model performance degrades when target (secondary) conditions have changed relative to the original calibration (primary) conditions. Accurate predictions occur when calibration and target samples are similarly affected by these inherent sample and measurement matrix effects. This matching constraint is the heart of restricting chemical analyses and limiting automatic calibration and prediction. Two frameworks are presented to alleviate the problems. One approach forms a calibration model orthogonal to the primary and secondary matrix effect differences followed by automatic selection of a collection of accurate predicting models from the thousands of models generated due to the multiple tuning parameters required. Unique to this model updating and selection approach is that no reference values for the secondary samples are used. The second method, local modeling, mines a library of thousands of spectra with respective reference analyte values for unique calibration sets composed of matrix effects equivalent to each target sample including analyte and interferent amounts. Current methods suffer because it is wrongly assumed that spectral similarity translates to matrix effect matching. The presented procedure solves this problem for automatic calibration and prediction. The two methods are applied to several near IR (NIR) datasets. Both approaches show substantial improvement over traditional prediction by a primary global model.

Automatic target-adaptive calibration and prediction frameworks

ACS Fall 2021

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