Ab initio dispersion potentials based on physics-based functional forms with machine learning.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning

Background:

• Accurate calculation of noncovalent interactions is crucial for understanding molecular behavior.
• Existing dispersion models like Grimme's D3 and many-body dispersion (MBD) have limitations in accuracy and applicability.
• Symmetry-adapted perturbation theory (SAPT) provides accurate interaction energies but is computationally expensive.

Purpose of the Study:

• Introduce SAPT10K, a large dataset of noncovalent interaction energies and components.
• Develop improved machine learning (ML) models for dispersion interactions.
• Enhance the accuracy and applicability of dispersion models for diverse molecular systems.

Main Methods:

• Computed 9982 interaction energies using SAPT2+(3)(CCD) with an aug-cc-pVTZ basis set, forming the SAPT10K dataset.
• Developed novel intermolecular features based on distance histograms for ML models.
• Trained ML models (D3-ML and MBD-ML) to predict dispersion energies, addressing deficiencies in existing models.

Main Results:

• The SAPT10K dataset covers diverse intermolecular complexes and potential energy surfaces.
• The developed D3-ML model, using only Cartesian coordinates, achieved superior performance on a test set compared to the CLIFF model.
• The ML models demonstrate applicability to a wide range of elements and charged monomers, surpassing limitations of other ML models.

Conclusions:

• SAPT10K provides valuable data for refining ab initio dispersion potentials.
• The new ML models offer a computationally efficient and accurate alternative for calculating dispersion interactions.
• These advancements facilitate the study of supramolecular assembly and chemical reactions by accurately modeling noncovalent forces.