CDS&E: Regularization Adaption Processes for Multivariate Calibration and Maintenance

点击翻译按钮获取中文摘要

In this project funded by the Chemical Measurement and Imaging program of the Chemistry Division, and the Computational and Data-Enabled Science and Engineering program of the Directorate for Mathematical and Physical Sciences, Professor John Kalivas of the Department of Chemistry at Idaho State University is developing new methods of multivariate calibration. Calibration is a multidisciplinary problem, but in analytical chemistry, it involves forming a quantitative mathematical relationship between an instrumental signal, such as from a hand held near infrared spectrometer, and a property of interest in a sample, such as the pulp content of trees or the amount of the active pharmaceutical ingredient in tablets. A key feature of the new mathematical calibration processes is the ability to include unlabeled data, for instance, spectra of samples without reference values. With unlabeled data, calibration costs and time will be substantially reduced. Another key feature is a mathematical process allowing adaption of a calibration to new measurement conditions with minimal effort, even if unlabeled data is not available. A minimum of four undergraduates from Idaho State University and a Hispanic professor as a Research Opportunity Award from Central New Mexico Community College in Albuquerque will learn state-of-the-art calibration methods and become proficient at performing scientific research including dissemination. Education of undergraduates in these advanced methods prepares them for subsequent scientific professional pursuits.This project focuses on new regularization processes for multivariate calibration and prediction based on the fundamentals of Tikhonov regularization. The calibration processes allow multiple penalties (tuning parameters) providing flexibility in forming accurate and precise calibrations and mechanisms to adapt a calibration to new conditions. With these regularization methods comes the ability to include unlabeled data, samples without reference values. Because the expensive component in calibration is obtaining reference values, costs in forming and maintaining calibrations can be substantially reduced with unlabeled data. In order to select optimal values for tuning parameters, new fusion ranking methods will be developed and applied. The overall objective is development of computational regularization processes to adapt a calibration to a new domain using multiple tuning parameters, multiple calibration merits, and unlabeled data if it is available. Results from this project will advance the multidisciplinary field of multivariate calibration. For example, process analytical technology for the pharmaceutical and chemical industries, environmental and agriculture monitoring, and medical diagnostics all rely on multivariate calibration. With the improvements being targeted in this project, these diverse fields may well be better equipped to perform and sustain calibrations. Another broader impact of this project is the planned dissemination of the calibration algorithms via the investigators' web sites, allowing newcomers and practitioners direct access.