Benchmarking chemical neural ordinary differential equations to obtain reaction network-constrained kinetic models from spectroscopic data

Abstract

Kinetic model identification relies on accurate concentration measurements and physical constraints to limit solution multiplicity. Not having these measurements and prior knowledge of species and reactions creates considerable challenges that are currently unresolved. We address these by developing a data-driven framework using realtime spectroscopic data, comprising: (i) multivariate curve resolution to deconvolve the spectra of the reacting mixture into those of its pseudocomponents and their corresponding concentrations, which enables species identification without prior information, (ii) Bayesian structure learning among the pseudocomponent spectra to enable hypothesizing reaction pathways, and (iii) neural ordinary differential equations (ODE) that are physically constrained by the hypothesized reaction network and the laws of mass action and temperature dependence to learn kinetic models from the temporal concentration projections of the realtime spectra. The predictive performance of the constrained neural ODEs is limited by the accuracy of spectral deconvolution in the presence of noise, and has been benchmarked against a constrained regression approach by varying signal:noise ratios in the synthetic spectroscopic data of reacting mixtures. Although the hypothesized reaction network differs from the actual reaction template, owing to noise, the network-constrained neural ODEs are seen to result in a 75.2% and 68.15% decrease in the root mean squared error (RMSE) of the concentration profile predictions as compared to the constrained regression method, when trained on time projected concentration data of the synthetic spectra generated at a signal to noise ratio of 35 and 100, respectively.