The exploration of reaction pathways lies at the heart of understanding chemical processes, with applications ranging from catalysis to drug design. A reaction path in its fundamental form describes the rearrangement of atoms between reactant and product state along a connected set of structures, providing insight into the kinetics of the reaction. With the advancement of computational hardware and quantum chemical methods in the recent years, computational and data-driven reaction network discovery has become accessible, revolutionizing the way chemical space is explored, thus accelerating the development of novel molecules and materials [1,2,3]. However, at the heart of this endeavour still lies the efficient generation of minimum energy paths (MEP), which are often needed a priori as feedstock for data driven methods. It has been shown that the MEP generated by common use gradient-based methods such as the Nudge Elastic Band (NEB)[4] or string[5] method heavily rely on the initial path and the coordinate space chosen[6]. In this regard we present a systematic and broad comparison of representations used in MEP generators, as well as a diverse dataset of benchmark reactions for reaction-exploration algorithms. Our dataset shows large chemical diversity, containing molecules with up to 21 heavy (non-hydrogen) atoms from the reaction space spanned by C, H,O, N, S, P, B, F, Cl and Br. By including larger and chemically more diverse molecules than previously established datasets [7,8] we provide an opportunity to test the robustness and predictive capabilities of established methodologies.
Maike Mücke