Learning (from) the Electron Density: Transferability, Conformational and Chemical Diversity.

Area of Science:

• Quantum Chemistry
• Machine Learning
• Computational Physics

Background:

• Machine learning is rapidly advancing in quantum chemistry, with applications to molecular properties.
• Electron density (ρ(r)) is crucial for density functional theory but challenging for ML models due to rotational symmetries.
• Existing ML models for ρ(r) face limitations in scalability and transferability.

Purpose of the Study:

• To develop a machine learning model for electron density (ρ(r)) that overcomes limitations in scalability and transferability.
• To accurately describe the covariance of electron density spherical tensor components.
• To demonstrate the model's predictive power and efficiency.

Main Methods:

• Developed a local regression framework using symmetry-adapted Gaussian process regression.
• Combined an efficient electron density decomposition scheme with the regression framework.
• Trained the model on local atomic environments for enhanced transferability.

Main Results:

• The symmetry-adapted Gaussian process regression model accurately describes electron density spherical tensor components.
• The model exhibits high transferability and linear-scaling prediction with the number of atoms.
• Demonstrated predictive power on various applications, including the electron density of Ubiquitin.

Conclusions:

• The proposed machine learning approach significantly improves the prediction of electron density (ρ(r)).
• The model offers enhanced scalability and transferability, crucial for large molecular systems.
• This method represents a significant advancement in applying machine learning to quantum chemistry.