Kohn-Sham accuracy from orbital-free density functional theory via Δ-machine learning.

Area of Science:

• Computational chemistry
• Materials science
• Machine learning

Background:

• Orbital-free density functional theory (DFT) offers computational efficiency but lacks accuracy.
• Kohn-Sham DFT provides higher accuracy but is computationally expensive.
• Bridging this accuracy-efficiency gap is crucial for large-scale simulations.

Purpose of the Study:

• To develop a machine learning model that achieves Kohn-Sham DFT accuracy using orbital-free DFT calculations.
• To improve the accuracy of energies and forces in orbital-free DFT.
• To enable efficient and accurate molecular dynamics simulations.

Main Methods:

• Implementation of a Delta-machine learning model using a kernel-based machine-learned force field (MLFF).
• Capturing the energy and force differences between Kohn-Sham and orbital-free DFT.
• On-the-fly molecular dynamics simulations for accuracy and performance assessment.

Main Results:

• The Delta-MLFF model significantly improves the accuracy of Thomas-Fermi-von Weizsäcker orbital-free energies and forces by over two orders of magnitude.
• The proposed model outperforms MLFFs based solely on Kohn-Sham DFT in accuracy.
• The model demonstrates greater efficiency and reduced sensitivity to parameters compared to existing methods.

Conclusions:

• The Delta-MLFF approach successfully bridges the accuracy gap between orbital-free and Kohn-Sham DFT.
• This method offers a computationally efficient pathway to accurate predictions in materials science.
• Application to molten Al0.88Si0.12 suggests no silicon aggregation, validating the model's predictive power.