Machine Learning Assisted Selective Configuration Interaction for Accurate Ground and Excited State Calculations.

Area of Science:

• Quantum chemistry
• Computational physics
• Machine learning applications

Background:

• Accurate electronic structure calculations are crucial for understanding molecular properties.
• Traditional methods like full configuration interaction (FCI) are computationally expensive.
• Selective Configuration Interaction (SCI) methods aim to reduce computational cost by selecting important electronic configurations.

Purpose of the Study:

• To introduce a novel perturbative Selective Configuration Interaction (SCI) approach.
• To integrate a machine learning classifier for efficient selection of Slater determinants.
• To achieve high accuracy in electronic structure calculations with reduced computational cost.

Main Methods:

• Development of a perturbative SCI approach guided by a binary machine learning classifier.
• Leveraging a lightweight feedforward neural network (FNN) for fast training.
• Utilizing a binary cross-entropy metric to identify important Slater determinants.

Main Results:

• The proposed method achieves accuracy comparable to the state-of-the-art SCI method CIPSI.
• The model reliably identifies key Slater determinants across various configuration-space sizes.
• Achieved FCI/CASCI-level accuracy within 10^-4 Hartree for both ground and excited states.
• Demonstrated robustness for strained geometries and conformational changes.

Conclusions:

• The machine learning-guided SCI approach offers a computationally efficient and accurate alternative for electronic structure calculations.
• This method successfully identifies crucial Slater determinants, enabling high-accuracy predictions.
• The approach shows promise for developing new regression-based strategies for molecular electronic structure and potential energy surfaces.