Neural Network Method for Constructing Intermolecular Potential Energy Surfaces of van der Waals Complexes

This study proposes a new approach for constructing intermolecular potential energy surfaces (PESs) of van der Waals (vdW) complexes using neural networks. The descriptors utilized in this neural network model are split into two parts: radial parts representing the intermolecular stretching vibrations between monomers and angular parts describing the relative orientation of these molecules. Specifically, the parity-adapted rotational basis functions used in the bound state calculation are taken as the angular descriptors, which ensure the correct symmetry of the PES. The number of orthogonal rotational basis functions is controlled by the maximum value of the angular momentum quantum number. In addition, the symmetry of monomer molecules is achieved by restricting the quantum number of the rotational basis function. The descriptors for five types of van der Waals complexes, including atom-linear, atom-nonlinear, linear-linear, linear-nonlinear and nonlinear-nonlinear molecules complexes, have been derived in this work. The neural network models with these newly developed descriptors were then applied to construct PESs of two van der Waals complexes, Ar-NaCl and N₂-OCS. The root-mean-square error values between the fitted and ab initio energies are found to be 0.11 cm⁻¹ and 0.26 cm⁻¹ for Ar-NaCl and N₂-OCS, respectively. These results indicate that this method is accurate and effective for constructing high-precision PESs of vdW complexes.