Fourth-Generation Neural Network Potentials for Molecular Chemistry

点击翻译按钮获取中文摘要

Machine learning potentials (MLP) have become an important tool for performing atomistic simulations of condensed systems with the accuracy of electronic structure methods at a small fraction of the computational costs. To date, most applications have been reported in materials science, while organic molecules have been primarily studied for benchmark purposes in vacuum. Although most chemical reactions occur in the liquid phase, applications of MLPs to solvation and molecular chemistry in solution are still very rare. Apart from the complexity of the involved configuration space, a major challenge for studying these systems is the need for a highly accurate description of intra- as well as intermolecular interactions, from strong covalent bonds via hydrogen bonding to electrostatic and dispersion interactions. A particularly crucial aspect is the charge distribution in the involved species, which cannot be captured correctly by most current MLPs based on local properties like environment-dependent atomic energies and charges.Recently, we have developed a fourth-generation high-dimensional neural network potential (4G-HDNNP), which combines the accurate description of local bonding and reactivity with long-range interactions based on the global charge distribution in the system. This global description is not only essential for molecules containing delocalized electrons, e.g. in aromatic groups or conjugated pi-systems, but also if the molecular charge is changing, like in (de)protonation, which is a key step in many types of reactions in organic chemistry. All these systems can in principle be studied by 4G-HDNNPs, which explicitly take into account the global charge distribution resulting from reactions, different functional groups and varying total charges, making this method a promising approach for molecular chemistry. The goal of this project is to explore the applicability of 4G-HDNNPs to molecular chemistry in solution by focusing on two major aspects, the quality of the density functional theory (DFT) reference data and the generalization of the 4G-HDNNP method. High-quality reference data will be obtained by benchmarking the reliability of exchange correlation functionals beyond the Generalized Gradient Approximation (GGA) level to Quantum Monte Carlo and Coupled Cluster calculations, and by including dispersion and self-interaction corrections (SIC). The 4G-HDNNP will be extended by employing new descriptor types for structural discrimination being applicable even to difficult situations like conical intersections and by the introduction of charge constraints, which, along with SIC and constrained DFT calculations, will allow to overcome the integer charge problem in both, DFT and the 4G-HDNNP, in a consistent approach. This new set of computational tools will be implemented in the open-source software RuNNer and applied to representative solute-solvent model systems covering important scenarios in synthetic organic chemistry.