Scientists seeking to understand and develop catalytic methods use a combination of different techniques to study their problems, including catalytic yield outcomes with different conditions or chemical compositions, kinetic experiments, in-situ observation of catalytic species and DFT calculations of transition states. Currently, machine learning and data analytics are becoming an increasingly crucial drivers for new catalyst development. While quite effective at predicting one type of output of complex processes in a black-box manner, standard ML approaches are unable to utilize connections between different data types to enhance understanding and predictivity, particularly with an eye towards the design of new catalytic methods. We have developed a joint kinetic-reactivity model, which allows us to fit yield data, kinetic traces and DFT-calculated transition states in one coherent model. This model consists of a linear or quadratic relation of chemical features of the substrates and catalysts involved with each rate in a simplified mechanism that can be directly connected to transition state calculations. The rates calculated for each reaction are then inserted into kinetic models to receive predictions of yields and kinetic traces. We apply this model to a well-researched homogeneous catalysis reaction, Pd-catalyzed C-N cross-coupling. Existing literature datasets of yields and DFT-calculated transition states are incorporated into the training data alongside new DFT calculations and kinetic traces. Next to applying this model to the problem of predicting yields for previously unreported catalyst and substrate combinations, we also showcase how the joint kinetic-reactivity model can be applied to predict behavior of catalytic reactions such as Hammet plots and how we can utilize this model to design new ligands for catalytic transformations.
