AP-Net: An atomic-pairwise neural network for smooth and transferable interaction potentials

Area of Science:

• Computational chemistry
• Materials science
• Chemical physics

Background:

• Accurate computation of intermolecular interactions is crucial but computationally expensive with ab initio methods.
• Existing machine learning (ML) potentials often lack physically realistic potential energy surfaces and exhibit deficiencies due to atomic energy partitioning.
• There is a need for ML models that provide accurate interaction energies and physically meaningful potential energy surfaces for intermolecular systems.

Purpose of the Study:

• To develop a novel machine learning framework and model (AP-Net) for accurate and physically interpretable computation of intermolecular interaction energies.
• To address the limitations of existing ML potentials, particularly concerning the smoothness and asymptotic behavior of potential energy surfaces.
• To demonstrate the transferability and accuracy of AP-Net across diverse chemical systems.

Main Methods:

• Introduction of a physically motivated atomic-pairwise framework for intermolecular ML potentials.
• Development of AP-Net, a neural network model leveraging the atomic-pairwise paradigm and symmetry adapted perturbation theory (SAPT) interpretability.
• Training and validation of AP-Net on datasets including hydrogen-bonded dimers and the S66x8 dataset.

Main Results:

• AP-Net generates smooth, physically meaningful intermolecular potentials with correct asymptotic behavior, unlike previous models.
• The model shows significant transferability, accurately predicting energies for the diverse S66x8 dataset after training on limited data.
• AP-Net achieved a mean absolute error of 0.37 kcal mol⁻¹ on experimental hydrogen-bonded dimers, with 2-5 fold error reduction in SAPT components.

Conclusions:

• The proposed atomic-pairwise framework and AP-Net model represent a significant advancement in ML potentials for intermolecular interactions.
• AP-Net successfully learns the underlying physics of intermolecular forces, including electrostatic interactions in hydrogen bonds.
• The model's accuracy, interpretability, and transferability offer a promising alternative to traditional computational methods for studying molecular interactions.