(376ad) Developing First-Principles Based Interatomic Potentials with Bayesian Inference

Fundamental understanding of the electronic and catalytic properties of metallic nanoparticles and surfaces has been widely studied using density functional theory (DFT) calculations over the last few decades. However, the high cost of computations has hindered employing first principles DFT calculations for a large number of atoms (over a thousand) in real-world catalyst systems. Many highly efficient methods of calculating the energetic and structural properties of atomistic systems have been developed to overcome the size limitation of the DFT method by employing interatomic potentials. One of the most commonly used potential formalism for metals is the embedded-atom method (EAM). In standard approaches, EAM potentials have been developed by fitting parameters of its functional to the data from experimental observations and/or theoretical calculations and single optimal values are determined for each parameter. The current fitting techniques do not represent the uncertainty about the accuracy of the developed interatomic potential which is the result of fluctuation in fitted parameter values. To attain an estimate of the errors, we took advantage of the Bayesian model calibration method. Bayesian learning of interatomic potentials from ab initio data while considering the impact of fluctuations of these parameters is promising as a general, robust approach for force field optimization. This can be used in studying the uncertainties of atomistic simulations of metallic systems such as the catalytic properties of Au nanoparticles, particularly at the grain boundaries that are important for electrocatalytic reactions. With the aid of EAM potentials developed here, we studied the structure-activity relationships of grain boundaries of gold NPs for CO₂ reduction. The accurate prediction of *CO and *COOH adsorption energies using the developed structure–reactivity relationships helped us in observation of reduced overpotential for CO₂ electroreduction reaction on grain boundaries region. This superior activity which can be attributed to the induced strain by dislocations was also captured experimentally.