(632c) Determining Polymer Size from Raman Spectroscopy

The size of polymer particles is a critical product quality and as such it is desired to monitor and control it during polymerization processes. Established methods for size determination, such as dynamic light scattering (DLS), require off-line measurements for high prediction accuracy. In contrast, in-line particle size determination is an open challenge. For in-line concentration monitoring, Raman spectroscopy constitutes a reliable process analytical tool. Theoretical considerations indicate that particles influence spectroscopic measurements by light scattering, and experimental evidence for the correlation between Raman scattering and polymer size has previously been presented [1-6]. State-of-the-art methods for predicting polymer sizes from Raman spectra rely on linear methods such as partial-least squares (PLS) or principal component analysis. However, these studies show a lack of prediction accuracy compared to the well-established DLS method.

We leverage nonlinear manifold learning approaches for improved prediction accuracy of particle size based on Raman spectroscopy. We propose three distinct machine learning workflows, all based on an initial dimensionality reduction of the Raman spectra via diffusion maps (DMAPs) [7-9]. The workflows encompass: direct prediction from diffusion map coordinates, prediction from alternating diffusion maps [10-12], and prediction via conformal y-shaped autoencoder neural networks [13]. We apply these workflows to a dataset comprising 47 Raman spectra from microgel (cross-linked polymer) samples coupled with size measurements obtained via DLS. In each sample, the microgels exhibit monodisperse diameter; between samples the diameter ranges from 208 nm to 483 nm. We compare the proposed workflows with the benchmark prediction approach via PLS.

The results indicate that among all the approaches studied, prediction via conformal autoencoders emerges as the most promising, outperforming state-of-the-art methods and achieving unprecedented prediction accuracy. We present for the first time an algorithmic workflow for predicting particle size from Raman measurements that is competitive with the accuracy of off-line DLS measurements. Therefore, the proposed workflows allow direct size prediction from in-line measurements taken from untreated analytes and without spectral pretreatment, thus avoiding labor-intensive off-line processing and enabling online reaction monitoring for closed-loop control. The high efficiency of the proposed workflow due to the algorithmic filtering via DMAPs is especially relevant for a proficient performance based on limited data availability, which is highly desirable in the context of laborious experimental procedures.

In the future, the proposed workflow can be applied to simultaneously predict polymer concentration and size from Raman spectroscopy, thus providing a comprehensive and powerful in-line analysis tool.

[1] M. M. Reis, P. H. Araújo, C. Sayer, and R. Giudici, “Evidences of correlation between polymer particle size and Raman scattering,” Polymer, vol. 44, no. 20, pp. 6123–6128, 2003.

[2] M. van den Brink, M. Pepers, and A. M. van Herk, “Raman spectroscopy of polymer latexes,” Journal of Raman Spectroscopy, vol. 33, no. 4, pp. 264–272, 2002.

[3] K. Ito, T. Kato, and T. Ona, “Non-destructive method for the quantification of the average particle diameter of latex as water-based emulsions by near-infrared Fourier transform Raman spectroscopy,” Journal of Raman Spectroscopy, vol. 33, no. 6, pp. 466–470, 2002.

[4] C. Houben, G. Nurumbetov, D. Haddleton, and A. A. Lapkin, “Feasibility of the simultaneous determination of monomer concentrations and particle size in emulsion polymerization using in situ Raman spectroscopy,” Industrial & engineering chemistry research, vol. 54, no. 51, pp. 12867–12876, 2015.

[5] P. M. N. Ambrogi, M. M. E. Colmán, and R. Giudici, “Miniemulsion polymerization monitoring using off-line Raman spectroscopy and in-line NIR spectroscopy,” Macromolecular Reaction Engineering, vol. 11, no. 4, p. 1600013, 2017.

[6] J. Meyer-Kirschner, A. Mitsos, and J. Viell, “Polymer particle sizing from Raman spectra by regression of hard model parameters,” Journal of Raman Spectroscopy, vol. 49, no. 8, pp. 1402–1411, 2018.

[7] R. R. Coifman and S. Lafon, “Diffusion maps,” Applied and Computational Harmonic Analysis, vol. 21, no. 1, pp. 5–30, 2006.

[8] B. Nadler, S. Lafon, R. R. Coifman, and I. G. Kevrekidis, “Diffusion maps, spectral clustering and reaction coordinates of dynamical systems,” Applied and Computational Harmonic Analysis, vol. 21, no. 1, pp. 113–127, 2006.

[9] R. R. Coifman, I. G. Kevrekidis, S. Lafon, M. Maggioni, and B. Nadler, “Diffusion maps, reduction coordinates, and low dimensional representation of stochastic systems,” Multiscale Modeling & Simulation, vol. 7, no. 2, pp. 842–864, 2008.

[10] R. R. Lederman and R. Talmon, “Learning the geometry of common latent variables using alternating-diffusion,” Applied and Computational Harmonic Analysis, vol. 44, 3, pp. 509–536, 2018.

[11] O. Katz, R. Talmon, Y.-L. Lo, and H.-T. Wu, “Alternating diffusion maps for multi-modal data fusion,” Information Fusion, vol. 45, pp. 346–360, 2019.

[12] F. Dietrich, O. Yair, R. Mulayoff, R. Talmon, and I. G. Kevrekidis, “Spectral discovery of jointly smooth features for multimodal data,” SIAM Journal on Mathematics of Data Science, vol. 4, no. 1, pp. 410–430, 2022.

[13] N. Evangelou, N. J. Wichrowski, G. A. Kevrekidis, F. Dietrich, M. Kooshkbaghi, S. Mc-Fann, and I. G. Kevrekidis, “On the parameter combinations that matter and on those that do not: data-driven studies of parameter (non)identifiability,” PNAS Nexus, vol. 1, pgac154, 09 2022.