Low-Data Deep Quantum Chemical Learning for Accurate MP2 and Coupled-Cluster Correlations.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning

Background:

• Accurate prediction of electronic energies for macromolecules using post-Hartree-Fock methods is computationally expensive.
• Developing efficient methods for calculating electron correlation energies is crucial for molecular modeling.

Purpose of the Study:

• To develop a data-efficient deep neural network (dNN) model for predicting machine-learned electron correlation energies.
• To achieve predictions on par with traditional methods like MP2 and CCSD for complex molecules.
• To demonstrate the transferability of the model across diverse molecular systems and datasets.

Main Methods:

• Exploiting physically justified local correlation features in a compact basis.
• Constructing an expressive deep neural network (dNN) model.
• Training the dNN model on small molecule datasets and testing on larger, unseen molecular systems.

Main Results:

• The dNN model achieved electron correlation energies comparable to MP2 and CCSD levels of theory.
• The model demonstrated high data efficiency and transferability across various molecular types, including alkanes, organic molecules, and water clusters.
• Accurate predictions for large water clusters ((H2O)128) were achieved by training on smaller clusters ((H2O)8), with errors within chemical accuracy.

Conclusions:

• A compact local correlation feature set, while insufficient for direct post-Hartree-Fock calculations, is highly effective for machine learning.
• The dNN-powered model offers a computationally efficient and accurate approach for predicting electron correlation energies.
• The findings highlight the potential of machine learning in advancing computational chemistry and molecular modeling.