Interpretable Machine Learning for Kinetic Rate Model Discovery

Abstract

Due to global scale problems (e.g. exponential population growth), catalytic processes are of more importance today than they have ever been before. The industrialisation of these processes requires kinetic models. Mechanistic models are difficult to construct; data-driven and hybrid models, although easier to construct, lack interpretability and physical knowledge. Recently, a new approach called automated knowledge discovery has been popularised, but existing methods in the literature suffer from important drawbacks: necessitating assumptions about model structures and a lack of model selection routine (both being directly linked to kinetic model building). As such, this motivated the presented work to construct a generalisable methodological framework for the automated discovery of catalytic kinetic models. The methodological framework proposed utilises symbolic regression for model generation, a hybrid optimisation algorithm for parameter estimation, and the Akaike information criterion (AIC) and the Hunter–Reiner criterion for model selection and discrimination, respectively. The methodology was applied to an illustrative isomerisation case study, where concentration versus time data were provided for three different experiments (i.e. each experiment has different initial conditions). The framework was able to retrieve the correct kinetic model with realistic (i.e. noisy) data from the catalytic system. This exemplifies how the presented methodology can be harnessed to efficiently provide important and interpretable insights of catalytical systems that have not yet been researched.