Transferable MP2-Based Machine Learning for Accurate Coupled-Cluster Energies.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning

Background:

• Machine learning (ML) models are increasingly used for molecular property prediction.
• Current ML models often rely on geometric molecular representations.
• Electronic structure offers a more detailed physical representation but presents unique challenges.

Purpose of the Study:

• To develop an efficient ML method for predicting high-cost coupled-cluster singles-and-doubles (CCSD) wave functions.
• To encode lower-cost correlated wave functions (e.g., from MP2) using electronic structure features.
• To explore novel molecular representations focusing on electron correlation.

Main Methods:

• Developed a new molecular representation based on correlation-pair energies and electron promotions.
• Differentiated models based on two-electron promotions from the same or different molecular orbitals.
• Engineered input features describing orbital properties relevant to electron correlation.

Main Results:

• The developed models are highly transferable and size-extensive.
• Achieved chemical accuracy with very few training instances across diverse organic molecules.
• Demonstrated efficiency and transferability on linear hydrocarbons, water dimer, and the GDB-9 database.

Conclusions:

• A novel electronic structure-based ML representation enables accurate and efficient prediction of molecular properties.
• The method requires minimal training data, showing excellent transferability.
• This approach significantly advances the low-cost evaluation of molecular properties using quantum mechanical data.