Semi-supervised machine learning approach for reaction stoichiometry and kinetic model identification using spectral data from flow reactors

Abstract

Continuous flow reactors integrated with spectroscopic instruments allow for rapid collection of informative spectral data. The measured spectral data and a calibration model can be used to monitor reaction progress, elucidate reaction kinetics, and gain mechanistic insights efficiently. However, developing a calibration model is a time and resource-consuming task. Here, we propose a novel calibration-free integrated model identification framework, called, a semi-supervised machine learning approach (SSML) for identifying reaction systems rapidly using spectral data with minimal labelled data. Using the proposed SSML approach, the stoichiometric matrix and physically meaningful extents of reaction are identified from spectral data alone without invoking kinetic models. Subsequently, the computed extents are used for kinetic model discrimination and parameter estimation using the incremental identification method. The proposed method is demonstrated using an enzymatic hydrolysis reaction and a complex Wittig reaction system carried out in a micro-reactor equipped with an in situ UV-visible spectrophotometer. The results from the proposed calibration-free modelling framework are compared with those obtained using the traditional calibration-based method.