Efficient Sampling for Machine Learning Electron Density and Its Response in Real Space.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Electron density is crucial for determining ground state electronic properties.
• Existing machine learning (ML) models for electron density require extensive basis functions or grid points.
• Developing efficient ML models for real-space electron density is challenging.

Purpose of the Study:

• To develop an efficient real-space grid-based ML model for electron density and its response to electric fields.
• To reduce the number of training points required for accurate electron density prediction.
• To enable analysis of charge distribution and electric field effects in various systems.

Main Methods:

• Implemented an efficient grid-point sampling strategy combining targeted sampling and feature screening.
• Integrated the sampling strategy with a field-induced recursively embedded atom neural network model.
• Applied the ML model to QM9 molecular data, H2O/Pt(111) interface, Au(100) electrode, and Au nanoparticle.

Main Results:

• Achieved comparably accurate predictions for electron density using significantly fewer training points than previous models.
• Successfully predicted electron density and its response to electric fields across diverse systems.
• Enabled accurate partial charge partitioning and analysis of charge variation in interfacial systems.

Conclusions:

• The proposed efficient grid-point sampling strategy significantly enhances ML models for electron density.
• The developed ML model accurately predicts electronic properties and responses to electric fields.
• This approach offers a computationally efficient method for studying electron density and charge dynamics.