Transferable Atomic Multipole Machine Learning Models for Small Organic Molecules.

Area of Science:

• Computational chemistry
• Quantum chemistry
• Machine learning applications

Background:

• Accurate molecular electrostatic potential representation is vital for understanding intermolecular interactions.
• Distributed multipole moments are commonly used for this representation.
• Efficient evaluation of these moments is computationally challenging.

Purpose of the Study:

• To develop a machine learning model for predicting multipole coefficients of common atom types (H, C, O, N, S, F, Cl).
• To enable accurate calculations of molecular electrostatic potential across diverse molecular conformations.
• To improve the efficiency of evaluating intermolecular interactions.

Main Methods:

• Training machine learning models on quantum-chemical data from thousands of organic molecules.
• Developing separate models for neutral, cationic, and anionic molecular charge states.
• Utilizing atom types H, C, O, N, S, F, and Cl in any molecular conformation.

Main Results:

• Demonstrated high predictive accuracy of the machine learning models for multipole coefficients.
• Successfully evaluated intermolecular interaction energies for nearly 1,000 molecular dimers.
• Calculated the cohesive energy of the benzene crystal, validating the model's applicability.

Conclusions:

• The developed machine learning models provide an accurate and efficient method for predicting molecular multipole moments.
• This approach significantly enhances the evaluation of intermolecular interactions and crystal energies.
• The models are applicable to various molecular charge states and conformations.