DeePKS + ABACUS as a Bridge between Expensive Quantum Mechanical Models and Machine Learning Potentials.

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Machine learning (ML) potentials enable accurate, large-scale molecular simulations.
• Training ML potentials requires extensive, computationally expensive quantum mechanical (QM) data.
• High-level QM methods like DFT and quantum Monte Carlo are costly to generate training data.

Purpose of the Study:

• To introduce Deep Kohn-Sham (DeePKS), an ML-based density functional theory (DFT) model.
• To demonstrate DeePKS's ability to reduce data requirements for ML potential training.
• To provide a computationally efficient method for generating high-accuracy QM data.

Main Methods:

• Developed DeePKS, an ML-based DFT model using a neural network functional.
• Trained DeePKS with a significantly smaller dataset compared to traditional ML potentials.
• Implemented the DeePKS scheme for periodic systems in the ABACUS DFT package.

Main Results:

• DeePKS achieves energies and forces comparable to high-level QM methods.
• Requires orders of magnitude less training data than conventional ML potentials.
• Successfully generates high-accuracy QM data for training ML potentials.

Conclusions:

• DeePKS effectively bridges the gap between expensive QM calculations and ML potentials.
• Enables the generation of large, accurate datasets for ML potential development.
• Offers an open-source, efficient solution for large-scale molecular simulations.