Kinetic analysis of multi-step reactions is a valuable tool for the mechanistic understanding of reaction systems. However, developing high-quality kinetic models involves several steps such as carrying out several carefully designed experiments, separation, isolation, characterization, proposing kinetic models, fitting them to measured data. The reaction rates observed from traditional batch reactors are apparent in nature, as changes in concentrations are often confounded by transport processes. On the other hand, continuous flow micro-reactors have superior heat and mass transfer rates, and it is possible to relate changes in concentrations to intrinsic kinetics. Traditional methods for inferring concentrations include chromatographic separations, which generally provide time-delayed information about the reaction system. Real-time, online or inline spectroscopic analysis makes it possible to relate the measured spectroscopic measurements to concentrations, provided a calibration model is available. Developing a calibration model requires labelled data, viz., spectroscopic measurements and corresponding chemical compositions and hence are time and resource-intensive. The developed calibration model applies only to the predefined experimental conditions and cannot be extrapolated. Moreover, significant effort is required to maintain the developed calibration model due to changes caused by uncontrolled disturbances and disturbances in process variables. Hence, it is important to develop calibration-free approaches to handle spectral data.
In this work, a new calibration-free semi-supervised machine learning approach is proposed for monitoring and kinetic modeling directly from the time evolved spectral data without prior knowledge of pure component spectra or kinetic models. The proposed method consists of two steps: (i) computing of physically meaningful extents of reaction from spectral data without imposing any kinetic model a priori, and (ii) identifying the kinetic model structure and parameters from a set of kinetic rate expressions using the computed extent of each reaction. The proposed approach is demonstrated using an enzymatic hydrolysis reaction in a micro-reactor integrated with an in-situ UV-visible spectrometer. The results from the proposed calibration-free semi-supervised machine learning approach are validated using traditional calibration-based methods.