Chembot: A Machine Learning Approach to Selective Configuration Interaction.

Area of Science:

• Quantum Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Accurate quantum chemistry calculations are crucial for understanding molecular behavior.
• Full configuration interaction (FCI) provides exact energies but is computationally prohibitive for most systems.
• Stochastic and selected configuration interaction methods offer approximations but can be computationally intensive or less accurate.

Purpose of the Study:

• To introduce Chembot, a novel machine learning approach for selective configuration interaction.
• To evaluate Chembot's efficiency and accuracy in obtaining FCI-quality energies.
• To compare Chembot's performance against established methods like Monte Carlo and heat-bath configuration interaction.

Main Methods:

• Development of Chembot using a support vector machine for configuration selection.
• Utilization of charge density matrix and configuration energy as key features.
• Implementation of heuristics to enhance training data quality.

Main Results:

• Chembot achieves near FCI-quality energies with significantly fewer iterations and determinants compared to Monte Carlo configuration interaction.
• Chembot demonstrates competitive or superior performance to heat-bath configuration interaction in terms of variational space size.
• The method's effectiveness is validated on challenging small molecular systems (H4, H2C, H2O).

Conclusions:

• Chembot represents a significant advancement in efficient and accurate quantum chemistry calculations.
• The machine learning approach offers a promising alternative for achieving high-accuracy electronic structure results.
• Further applications of Chembot are expected to accelerate research in computational chemistry.