Predicting the Electron Density of Charged Systems Using Machine Learning.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Predicting electron density is crucial for calculating molecular and crystal properties using first-principles methods.
• Machine learning (ML) models can accelerate these computations by predicting electron density from atomic structure alone.
• Existing ML models struggle to predict electron density for charged systems.

Purpose of the Study:

• To extend the DeepDFT ML model to accurately predict electron density for charged systems.
• To introduce a novel input charge representation for charged systems.
• To validate the enhanced model across diverse charged material types.

Main Methods:

• Modified the DeepDFT ML model to incorporate system charge as an input feature.
• Developed and implemented a new input charge representation technique.
• Tested the model on various charged systems including defective perovskites, LiCoO2 supercells, diamond defects, MOFs, and molecular crystals.

Main Results:

• The enhanced DeepDFT model successfully predicts electron density for charged systems.
• The input charge representation approach proved effective across multiple test cases.
• Achieved accurate predictions for diverse charged materials, demonstrating model versatility.

Conclusions:

• The modified DeepDFT model provides an accurate and efficient method for predicting electron density in charged systems.
• This work overcomes a significant limitation in ML-based electron density prediction.
• The approach has broad applicability in computational materials science and chemistry.