Transferable Machine-Learning Model of the Electron Density.

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Electronic charge density is fundamental to atomic-scale material properties.
• Accurate computation of charge density typically requires intensive electronic-structure calculations.
• Developing efficient methods to predict charge density is crucial for advancing computational chemistry.

Purpose of the Study:

• To introduce a novel machine-learning framework for predicting valence charge density.
• To develop a transferable model capable of learning from small molecules and predicting for larger systems.
• To achieve low, linear-scaling computational cost for charge density predictions.

Main Methods:

• An atom-centered, symmetry-adapted framework was developed for machine learning.
• The model was trained on a limited set of reference electronic-structure calculations.
• Transferability was tested by training on small hydrocarbons (butane, butadiene) and predicting for larger ones (octane, octatetraene).

Main Results:

• The machine-learning model accurately predicts valence charge density.
• The model demonstrates high transferability across hydrocarbon molecules of varying complexity.
• Predictions for larger molecules like octane and octatetraene were accurate after training on smaller molecules.

Conclusions:

• A data-driven, transferable model for predicting electronic charge density has been established.
• This approach significantly accelerates electronic structure calculations.
• The model facilitates the computation of electrostatic interactions and aids in experimental interpretation.