(340bl) Development of in Situ monitoring and Data-Driven Modeling for Complex Systems: Case Study on Simulant Mixtures of Nuclear Waste

Authors: 
Kocevska, S. - Presenter, Georgia Institute of Technology
Maggioni, G. M., Bayer AG
Grover, M., Georgia Tech
Rousseau, R., Georgia Institute of Technology
Research Interests: Process Analytical Technology, Machine Learning, Spectroscopy, Reaction Kinetics

Approximately 56 million gallons of nuclear and chemical waste are stored at Hanford in Washington State. The clean-up of the site, which includes stabilizing the waste into a glass form via vitrification, is expected to continue for many years. One of the main challenges is the complexity of the waste, which calls for extensive sampling during processing. To address this challenge, in-line monitoring will be used to facilitate the continuous operation of the melter, while helping minimize employee exposures to hazards. My research explores the use of process analytical technology (PAT), such as infrared (IR) and Raman spectroscopy, coupled with machine learning methods, to measure and quantify the waste composition in real time.

As part of this work, we developed a flexible data-driven modeling framework to quantify the concentration of complex mixtures using spectroscopy data. The signal-to-composition model is based on our knowledge of the system and can vary from simple regressions for well-defined mixtures to advanced signal separation techniques, such as blind source separation (BSS), if we lack information on the waste components. One example of a complex mixture system is the direct-feed low-activity waste (DFLAW) feed, which has many constituents. However, only the concentrations of a few target species may need to be quantified in real-time to facilitate operation of the clean-up process. To obtain accurate estimates of the concentrations of the target species, without performing lengthy calibrations, we propose a four-step procedure that combines BSS techniques, with a traditional multivariate regression, such as partial least-squares regression (PLSR).

The BSS-PLSR procedure uses BSS to preprocess the data by (1) identifying the independent components in the mixture, (2) correlating the independent components with reference spectra to classify them as either target species (part of the critical quality attributes that need to be measured during waste processing) or non-target species, and (3) subtracting the contributions of non-target species. Next, the procedure uses PLSR for (4) quantification of the target concentrations in the BSS-preprocessed mixture spectra. Our research shows that incorporating BSS-preprocessing to remove non-target species can reduce the size of the PLSR training data set, while providing accurate estimates of the target species. The approach is tested for Raman and IR spectroscopy using simulated and experimental data sets based on simulant mixtures of nuclear waste. However, the developed algorithm is not restricted to Raman and IR, but in principle is applicable to any quantitative spectroscopic technique complying with physico-mathematical assumptions, such as the Beer-Lambert law.

Checkout

This paper has an Extended Abstract file available; you must purchase the conference proceedings to access it.

Checkout

Do you already own this?

Pricing


Individuals

AIChE Members $150.00
AIChE Emeritus Members $105.00
AIChE Graduate Student Members Free
AIChE Undergraduate Student Members Free
Non-Members $225.00