Machine learning interatomic potentials for the dynamics of transition metal catalysis

Current global challenges, such as the energy and climate crises, elicit a constant demand for new materials with specific properties required for driving the development of novel technologies. Transition metal complexes (TMCs) in particular, are interesting since they can be utilized for a wide range of diverse applications, for example, as catalysts for facilitating the hydrogenation reaction of carbon dioxide to green methanol. Despite recent advances, traditional approaches such as density functional theory and ab initio molecular dynamics to study their properties often suffer from either prohibitively high computational costs or insufficient accuracy. In recent years, machine learning interatomic potentials (MLIPs) have emerged as a promising alternative for studying complex and dynamical chemical phenomena in both gas and condensed phases. While there has been a lot of progress in the development and application of MLIPs, their extension to condensed phase reactivity, especially in systems containing metals, has remained largely unexplored.This fellowship aims to overcome the current limitations of the field by introducing an MLIP framework suitable for the study of condensed phase reactivity of TMCs. To achieve this goal, we will evaluate the performance of different MLIP frameworks including various atomic descriptors as well as neural network architectures. Furthermore, we will investigate transfer learning approaches that aim to leverage data of similar systems in order to reduce the overall amount of required training data. Training of the MLIPs will be facilitated through active learning strategies that efficiently sample training data points on the fly. A simple derivate of a Ruthenium MACHO catalyst for the hydrogenation reaction of carbon dioxide will serve as the initial toy model to test and benchmark the best strategies. Ultimately, the devised methodology will be applied to more complex systems with concrete real-world applications.