Machine learning prediction of dispersion interactions using the exchange hole dipole moment model

Abstract

Newly developed Neural Network Potentials (NNPs) like ANI are powerful tools to describe these systems with a high level of accuracy but with a modest computational cost. Because they use short-range cutoffs (e.g., 5 Å), interactions outside this range are not described correctly. Significantly, London dispersion is a fundamental intermolecular force that extends beyond 5 Å, so it is neglected by current NNPs. The dispersion interaction in a chemical system can be estimated as the sum of pairwise atomic interactions with 6th, 8th, and 10th order terms (i.e., C₆, C₈, and C₁₀). The exchange hole dipole moment (XDM) model provides an accurate ab initio method for calculating these coefficients. However, a computationally intensive Density-Functional Theory (DFT) calculation is required to calculate these coefficients, so they cannot be used practically with NNPs. In this thesis, we developed a neural network to calculate these dispersion coefficients of atoms without the DFT calculation, providing an NNP that describes dispersion rigorously while maintaining the computational efficiency of NNPs. This new NNP is trained to reproduce PBE0/aug-cc-pVTZ with an XDM dispersion correction. This method was validated by comparison to high-level ab initio calculations from the DES15K test set. This method predicted intermolecular interaction energies of neutral molecules with a mean absolute error of 1.1 kcal/mol and a coefficient of determination of 0.91, demonstrating that this method has comparable accuracy to the QM method with a dramatically reduced computational cost.