Self-learning path integral hybrid Monte Carlo with mixed ab initio and machine learning potentials for modeling...

Area of Science:

• Computational materials science
• Quantum mechanics
• Machine learning applications

Background:

• Machine learned potentials (MLPs) offer ab initio accuracy at lower computational cost for quantum simulations.
• Developing MLPs requires large, diverse ab initio training datasets, which are computationally expensive to generate.
• Efficiently exploring the phase space for training data is crucial but challenging.

Purpose of the Study:

• To present a novel self-learning path integral hybrid Monte Carlo method using mixed ab initio and ML potentials (SL-PIHMC-MIX).
• To enable the study of larger systems and extend existing self-learning methods to path integral simulations.
• To demonstrate the method's efficiency in reproducing ab initio results with fewer computations.

Main Methods:

• Developed the self-learning path integral hybrid Monte Carlo method with mixed potentials (SL-PIHMC-MIX).
• Utilized a combination of ab initio and ML potentials for enhanced computational efficiency.
• Applied the method to generate MLPs and perform path integral simulations.

Main Results:

• The SL-PIHMC-MIX method allows for the study of larger systems than previously possible.
• Simulations using SL-PIHMC-MIX with trained MLPs accurately reproduce structures obtained from ab initio path integral molecular dynamics (PIMD).
• Achieved exact reproduction of ab initio PIMD structures using only 5000 evaluations, compared to 100,000 for ab initio PIMD.

Conclusions:

• SL-PIHMC-MIX provides a computationally efficient pathway to achieve high-accuracy simulations of nuclear quantum effects.
• The method overcomes the challenge of generating large ab initio training datasets.
• Enables accurate and efficient exploration of complex systems where nuclear quantum effects are significant.