A Machine Learning Approach for MP2 Correlation Energies and Its Application to Organic Compounds

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning Applications

Background:

• Accurate simulation of chemical compounds requires proper treatment of electron correlation effects.
• Traditional post-Hartree-Fock methods for calculating correlation energies are computationally expensive and limited in scale.

Purpose of the Study:

• To develop a machine learning-based density functional approximation for efficient calculation of electron correlation energies.
• To achieve results comparable to second-order Møller-Plesset perturbation (MP2) calculations with reduced computational cost.

Main Methods:

• Utilized neural networks to create a novel density functional approximation.
• Trained the model on a small dataset of molecules at equilibrium structures.
• Evaluated transferability across different basis sets, structures, and nuclear configurations.

Main Results:

• Achieved results comparable to MP2 calculations for organic compounds.
• Demonstrated high accuracy with generally less than 5% relative error for out-of-training domain structures and systems up to ~140 atoms.
• Successfully applied the approach to predict correlation energies from molecular dynamics trajectories using minimal training data.

Conclusions:

• The proposed machine learning approach offers a computationally efficient alternative for calculating electron correlation energies.
• This method shows excellent transferability and accuracy, making it suitable for large-scale simulations and dynamic processes.