Understanding the mechanism of catalytic reactions is essential for improving the development of more efficient and sustainable catalysts. One of the largest pieces of this understanding comes from kinetic analysis, which evaluates the consumption of reactants and the formation of products over time. In this analysis, rates of product formation and reactant consumption are described by equations known as rate laws. While rate laws may be theoretically straightforward, in practice, their derivation is complex and requires many mathematical approximations that can lead to errors and that are also limited to reaction networks with only a few steps operating under steady state. In a recent article, Jordi Burés and Igor Larrosa introduced a method to streamline the study of reaction mechanisms using a deep neural network model that classifies the mechanisms of catalytic reactions based on the time-course (kinetic) signatures of the reactions, eliminating the need for rate law derivations and extending the analysis to non-steady-state kinetics.
The proposed model contains two types of neural networks: a long short-term memory network for processing sequences of temporal data, and a fully connected neural network to process atemporal data such as initial concentrations of the catalyst. The authors considered 20 reaction mechanisms and described each of them using a set of ordinary differential equations (ODEs). By solving these ODEs, the authors were able to generate 5 million kinetic samples for the training and validation of the model. Such an approach was particularly helpful because it avoided the bottleneck of producing the large amount of kinetic data needed to train deep learning models, which is especially difficult and time-consuming in a laboratory setting. The chosen reactions belonged to four categories: 1) the core Michaelis–Menten-type mechanism, 2) mechanisms with bicatalytic steps involving either catalyst dimerization or the reaction between two catalytic species, 3) mechanisms with catalyst activation steps, and 4) mechanisms with catalyst deactivation steps. The trained model was evaluated on 100,000 kinetic samples (5,000 samples per mechanism), in which it correctly categorized the profiles to the correct mechanistic class with 92.6% accuracy. Interestingly, even when noisy data was introduced, the accuracy remained high. The proposed model was also benchmarked against previously reported experimental kinetic profiles, in which the predicted mechanisms agreed with the existing studies and even provided additional mechanistic details that were overlooked in some of the original work, such as the product of a deactivation pathway. This work automates an arduous process and expands the ability to use kinetic analysis in reaction mechanism investigation.
This is a preview of subscription content, access via your institution