Charge-Optimized Electrostatic Interaction Atom-Centered Neural Network Algorithm.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Machine-learning potentials commonly use partial charges to model electrostatic interactions.
• Existing methods often require pretrained models and reference charges, adding complexity and limiting accuracy.
• Charge partition methods can be system-dependent, affecting model reliability.

Purpose of the Study:

• To develop a self-consistent machine-learning algorithm for predicting atomic energy and charges without reference data.
• To integrate these atomic charges into force field models for accurate electrostatic interaction calculations.
• To evaluate the algorithm's performance on diverse benchmark systems.

Main Methods:

• An atom-centered neural network (ANN) algorithm was developed, requiring only one model per element.
• The ANN predicts atomic energy and partial charges, which are used for electrostatic energy computation via Ewald summation.
• Force field models were trained using total energy, forces, and electrostatic energy.

Main Results:

• The ANN algorithm demonstrated reasonably accurate predictions of partial charges and electrostatic interactions.
• The method was tested on Ge slab, TiO2 crystalline, and Pd-O nanoparticle systems.
• The approach provides a self-consistent charge prediction strategy.

Conclusions:

• The proposed ANN algorithm offers a robust and reliable method for modeling electrostatic interactions in machine-learning potentials.
• This approach simplifies force field development by removing the dependency on reference charges.
• The self-consistent charge prediction enhances the applicability of machine learning in materials science.