Δ-Machine Learned Potential Energy Surfaces and Force Fields.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning

Background:

• Machine-learned potential energy surfaces (PESs) are advancing for larger molecular systems.
• Current methods often limit electronic structure theory to below coupled cluster with singles and doubles (CCSD(T)) accuracy.
• Existing datasets like MD17 use Density Functional Theory (DFT) (PBE) for energies and forces.

Purpose of the Study:

• To present and demonstrate a Δ-machine learning method for achieving near CCSD(T) accuracy in DFT-based PESs.
• To extend the Δ-machine learning approach to improve force fields.
• To develop accurate, ab initio potentials for molecular systems.

Main Methods:

• Application of a novel Δ-machine learning method to DFT-based PESs.
• Extension of Δ-machine learning to force fields using many-body corrections.
• Development of CCSD(T) datasets for 2-body, 3-body, and 4-body interactions in water.

Main Results:

• Achieved near CCSD(T) accuracy for DFT-based PESs of hydronium, N-methylacetamide, acetyl acetone, and ethanol.
• Demonstrated the extension of Δ-machine learning to force fields.
• Developed a new fully ab initio water potential, q-AQUA, using CCSD(T) interaction datasets.

Conclusions:

• Δ-machine learning is effective in elevating DFT-based PESs to near CCSD(T) accuracy.
• The method shows promise for improving force fields and developing highly accurate molecular potentials.
• Specialized techniques may be required for very large systems like 15-atom tropolone to reach CCSD(T) energies.