A Deep Learning-Augmented Density Functional Framework for Reaction Modeling with Chemical Accuracy.

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning

Background:

• Accurate prediction of reaction energetics is crucial but challenging for conventional computational chemistry methods.
• Density functional theory (DFT) often compromises accuracy for efficiency.
• High-level quantum mechanical methods provide accuracy but are computationally expensive.

Purpose of the Study:

• Introduce Deep post-Hartree-Fock (DeePHF), a novel machine learning framework.
• To achieve coupled cluster with single, double, and perturbative triples (CCSD(T))-level accuracy in reaction energetics prediction.
• To maintain the computational efficiency characteristic of DFT.

Main Methods:

• Integrate neural networks with quantum mechanical descriptors.
• Establish a direct mapping between local density matrix eigenvalues and high-level correlation energies.
• Develop a machine learning model trained on small-molecule reaction data.

Main Results:

• DeePHF achieves CCSD(T)-level precision in predicting reaction energetics.
• The framework demonstrates superior performance and exceptional transferability across benchmark datasets.
• Maintains O-(N^3) scaling, offering significant computational efficiency.
• Outperforms advanced double-hybrid functionals in accuracy.

Conclusions:

• DeePHF effectively bridges the gap between high-accuracy quantum chemistry and scalable computational models.
• The model circumvents the traditional accuracy-scalability tradeoff in computational chemistry.
• DeePHF presents a promising advancement for chemical reaction modeling.