(263c) Application of Blind Source Separation and Partial Least-Squares Regression to Quantify Target Species in Complex Nuclear Waste Mixtures
As part of our research, process analytical technology (PAT), such as in situ Raman and infrared spectroscopy, have been developed to measure the waste composition in real time. In this work we present a flexible data-driven modeling framework, in which we develop a signal-to-composition model based on knowledge of the liquid low-activity waste (LAW) system. The LAW contains many species, out of which nitrate, nitrite and sulfate need to be quantified in real time to facilitate continuous processing. In the usual approaches, a chemometrics model, such as partial least-squares regression (PLSR) is developed and trained with all the species in the waste. However, such an approach requires extensive model calibration involving both the target species (such as nitrate, nitrite and sulfate), and the many other species that may not need to be quantified. In order to reduce the number of experiments required to build a PLSR model, we propose a new methodology that couples PLSR with blind source separation (BSS). BSS methods, such as independent component analysis (ICA), use little or no prior information to decompose mixture data into its constituents, relying on minimization of the mutual information or maximization of the non-Gaussianity of the data. 
In our work, BSS is used as a preprocessing step, to identify and remove the signals of the non-target species. The blind source separation (BSS) â partial least-squares regression (PLSR) is carried out in four steps:
- Apply BSS methods, such as independent component analysis (ICA) and multiple curve resolution â alternating least-squares (MCR-ALS), which use the mixture spectra to identify a sources and mixing matrices, to identify the independent components (sources) in the waste mixtures;
- Compare the independent component (sources) spectra to a library of reference spectra of the target species using a matrix of correlation coefficients. Based on the correlation, classify the independent sources in two categories: target species (which correlate with the library) and non-target species (which do not correlate with the library);
- Calculate and remove the signals of the non-target species from the mixture spectra;
- Develop a partial least-squares regression model to quantify the targets. Since the signals of the non-target species have been removed, the model can be trained with only the targets, reducing the need for extensive calibration.
The BSS-PLSR framework was tested on Raman and infrared spectroscopy data using simulated and experimental measurements. Preprocessing the data with BSS resulted in greater accuracy in the concentration predictions for both spectroscopic techniques. One of the main advantages of using BSS to preprocess the data is removing the dependence of the PLSR model on the non-target species. Therefore, even if the number and concentrations of non-target species fluctuates throughout the process, the training data set for the PLSR does not need to be updated. While the BSS-PLSR framework was tested on Raman and IR spectra, it can theoretically be applied to any other spectroscopic technique.
 Peterson, R. A.; Buck, E. C.; Chun, J.; Daniel, R. C.; Herting, D. L.; Ilton, E. S.; Lumetta, G. J.; Clark, S. B. Review of the Scientific Understanding of Radioactive Waste at the U.S. DOE Hanford Site. Environmental Science & Technology 2018, 52, 108381â396
 Poirier MR, Howe AM, Miera FR, Stone ME, DiPrete CC, Farrar ME. WTP Real-Time In-Line Monitoring Program Tasks 4 and 6: Data Quality and Management and Preliminary Analysis Plan. 2017.
 HyvÃ¤rinen, A. & Oja, E. Independent component analysis: algorithms and applications. Neural Networks 13, 411â430 (2000).