A Scalable and Generalizable Method to Minimize Solvent Interference in Identification of Chemical Reaction Networks from Spectroscopic Data.

Challenges such as varying levels of solvent interference that obscure spectral bands restrict the applicability and direct adoption of spectroscopic techniques for the analysis and characterization of complex reacting systems. In this work, we develop a generic and scalable method to minimize solvent interference on the spectroscopic signatures of reacting mixtures under varying process conditions without prior information about the constituents. The method frames solvent effect minimization as a tensorial factorization problem to segregate the solute and solvent contributions (i.e., latent factors) across each data dimension. We employ two distinct methodologies, named the direct and orthogonal approaches, to distinguish between the solute and the solvent latent factors. Comparative analyses on four case studies with spectroscopic process data show the efficiency of the proposed methods in minimizing and extracting useful information from obscured bands. The extracted solvent-free latent factors can be reconstructed to provide solvent-free spectroscopic data or directly applied to tasks such as mixture characterization, impurity detection, predictive modeling, and data mining. In this work, we apply them to generate plausible reaction networks for various chemical systems. The proposed approaches generalize to any solvent and adapt to the large process data sets typically found in chemical process industries.