CDS&E: Adaptive Learning for Multivariate Calibration with Big Data Attributes

点击翻译按钮获取中文摘要

With support from the Chemical Measurement and Imaging program in the Chemistry Division, and partial co-funding from the Office of Investigative and Forensic Sciences in the National Institute of Justice, Professor John Kalivas and his group at Idaho State University are developing new methods of data analysis with the target of solving complex calibration problems. Calibration is a multidisciplinary problem. In chemistry it involves forming a mathematical relationship between measured electronic signals derived from instrumental measurements and information of interest in a sample, such as the nitrogen content of plant leaves, percent fat in beef, or the blood glucose level. With the growth of the amount of generated and shared data, there is a need for a paradigm shift in calibration methods. The Kalivas group is devising means to exploit historical calibration data to improve the utility of new chemical measurements. A strategic feature of the approach is the development of new mathematical tools enabling adaptation of field-based measurements to new conditions and sample types without complex laboratory analyses, thereby reducing analysis time and costs. Through participation in this work, undergraduates from Idaho State University are learning state-of-the-art calibration methods and becoming proficient at scientific research, including dissemination. Because of the multidisciplinary nature of the project, outcomes directly benefit industry and society with efficient algorithms for rapid and accurate analysis of samples. This project exploits the increasing availability of spectral databases to develop completely new computational processes and accompany algorithms to address the growing need for improved and simplified multivariate calibration. The fundamentals of the inherent chemical and physical molecular interactions responsible for the measured signal are considered along with instrument-specific issues (conditions) to create new self- and cross-modeling data mining tools. Unlike conventional global and local modeling, the method does not require optimization of any tuning parameters. Additionally, reference values are not needed for target sample conditions; the method is therefore considered to be semi-supervised learning. A key goal of the project is to advance multivariate calibration to a "big-data" level, developing efficient algorithms to reveal useful information. Using reference databases, the data-mining algorithms identify samples best matrix-matched to new samples (one or many). Results from this project will advance multivariate calibration in a range of applications, including process analytical technology for the pharmaceutical and chemical industries, environmental and agriculture monitoring, and medical diagnostics. With the improvements contributed by this project, these fields will be better able to form and sustain calibrations on site, removing the need for complex chemical analysis. All developed algorithms will be posted to the Kalivas web site, allowing free access to potential users.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.