Approaches to the Interpretation of Machine Learning Models Trained with Big Experimental Kinetic Data: An Example of the Suzuki–Miyaura Reaction

This paper presents the results of the development and interpretation of machine learning models using big experimental datasets where different ways are implemented to take into account information on the kinetics of reaction using the Suzuki–Miyaura reaction as an example. An original method for interpreting the resulting models was proposed, which consists in the multistep ablation of predictors used in the models, allowing for the identification of the most important reaction parameters required for the successful operation of catalytic systems. The obtained sets of the parameters determining the success of the reaction (rate, selectivity, and yield of the target product; E-factor and other environmental metrics of the reaction, energy efficiency, cost, etc.) as the minimal sets of the predictors of ML models that describe satisfactorily the differential and/or integral quantitative reaction parameters allowed us for the first time to identify the patterns of the functioning mechanism of catalytic systems in the Suzuki–Miyaura reaction, consistent with big experimental kinetic data obtained in a wide range of reaction conditions (nature and amounts of substrates, reagents, catalyst precursors, bases, additives, and solvents; temperature, stirring speed, and type of an atmosphere in the reactor).