Insights into one-body density matrices using deep learning.

Area of Science:

• Condensed Matter Physics
• Quantum Chemistry
• Computational Science

Background:

• The one-body reduced density matrix (1-RDM) is crucial for understanding many-body quantum systems.
• Accurate approximations for the 1-RDM as a functional of charge density are lacking.
• Deep learning is emerging as a powerful tool for developing accurate density functionals.

Purpose of the Study:

• To develop improved approximations for the 1-RDM using machine learning.
• To investigate the constraints and fundamental features of the 1-RDM through data distillation.
• To integrate physical knowledge into deep learning models for enhanced accuracy.

Main Methods:

• Training autoencoders on a dataset of 1-RDMs from exactly solvable models.
• Applying principal component analysis to understand data compressibility and constraints.
• Feature engineering by incorporating known physical properties of the 1-RDM.
• Comparing various deep learning architectures and approaches.

Main Results:

• Machine learning successfully identifies key constraints within the 1-RDM data.
• Learned constraints inform the development of better approximations for the 1-RDM.
• Integration of physical knowledge improves the performance of deep learning models.
• Insight gained into which physical features of the 1-RDM are most amenable to machine learning.

Conclusions:

• Deep learning, particularly autoencoders, offers a promising avenue for approximating the 1-RDM.
• Combining machine learning with physical insights leads to more accurate and interpretable models.
• This approach advances the development of density functional theory and the study of quantum systems.