(411g) Study of Density Functional Errors in Microkinetic Modeling: A Data Driven Approach

Due to their tractable computational cost and relatively high accuracy, Density Functional Theory (DFT) based methods are widely used in computational chemistry to calculate the ground state electronic properties of molecules. Specifically, DFT calculated electronic energies combined with statistical mechanical temperature corrections are used to parametrize microkinetic models of reaction pathways. These ab-initio parametrizations involve calculating enthalpies and free energies of molecules in the gas or adsorbed phase. While such models have had some notable successes (for example in identifying dominant reaction pathways), they suffer from highly uncertain rate predictions. A big part of their uncertainty stems from the high sensitivity of rates to uncertainty in ab-initio calculated energies. Hence, improving the accuracy of these energies would lead to a substantial improvement in the predictive performance of such hierarchical models. However, it is well known that errors are introduced in DFT calculated properties because of the choice of approximate functionals. These approximations trade accuracy for speed, making the calculation of practical molecules computationally tractable. One way to improve ab-initio energy parameters is to introduce systematic corrections to species in the model. These are done to off-set the DFT error. Many such ad-hoc methods have been proposed in the literature. However, a systematic framework to introduce such corrections is lacking. In this study, we investigate the errors introduced by commonly used DFT functionals in calculating gas phase properties. Using data-driven techniques we develop a framework to interpret these errors and assign them to easily observable properties of molecules. We thus develop an empirical understanding of systematic errors introduced by DFT calculations in the context of catalytically relevant chemistries. These corrections can easily be applied to specific molecules in the model, improving the overall accuracy of predictions.