Quantum chemical accuracy from density functional approximations via machine learning.

Area of Science:

• Computational Chemistry
• Quantum Chemistry
• Machine Learning Applications

Background:

• Kohn-Sham density functional theory (DFT) is widely used but often lacks the accuracy required for precise chemical predictions.
• High-accuracy ab initio methods like coupled-cluster are computationally expensive, limiting their use to small systems.
• Existing DFT functionals typically yield accuracies of 2-3 kcal·mol⁻¹.

Purpose of the Study:

• To develop a machine learning approach for calculating accurate coupled-cluster energies using DFT densities.
• To investigate the efficiency of density-based Δ-learning (Δ-DFT) in reducing training data requirements.
• To demonstrate the application of Δ-DFT for enhancing molecular dynamics (MD) simulations.

Main Methods:

• Leveraging machine learning to predict coupled-cluster energies directly from DFT-calculated electron densities.
• Implementing density-based Δ-learning (Δ-DFT) to learn corrections to standard DFT calculations.
• Incorporating molecular symmetries to further reduce the necessary training data.
• Applying Δ-DFT to correct "on the fly" DFT-based MD simulations of resorcinol.

Main Results:

• Achieved quantum chemical accuracy (errors < 1 kcal·mol⁻¹) on test data by predicting coupled-cluster energies from DFT densities.
• Demonstrated that Δ-DFT significantly reduces the amount of training data needed, especially with symmetry considerations.
• Successfully obtained MD trajectories with coupled-cluster accuracy for resorcinol by correcting DFT simulations using Δ-DFT.
• Showcased the robustness of Δ-DFT in handling strained geometries and conformer changes.

Conclusions:

• Δ-DFT enables machine learning-driven calculations of coupled-cluster quality energies from DFT densities.
• The Δ-DFT approach significantly reduces data requirements, making high-accuracy computations more feasible.
• This method facilitates gas-phase MD simulations with quantum chemical accuracy, overcoming limitations of standard DFT for complex molecular systems.