Accelerating Ab Initio Simulation via Nested Monte Carlo and Machine Learned Reference Potentials.

Area of Science:

• Computational materials science
• Atomic-scale modeling
• Quantum chemistry

Background:

• Ab initio simulation is crucial for atomic-scale materials modeling.
• Ab initio Monte Carlo (AIMC) and ab initio molecular dynamics (AIMD) are key simulation strategies.
• AIMC offers exact thermodynamic sampling, while AIMD provides computational efficiency.

Purpose of the Study:

• To enhance the performance of AIMC simulations.
• To achieve efficiency levels comparable to or exceeding AIMD.
• To demonstrate a novel hybrid simulation strategy.

Main Methods:

• Developed a hybrid nested sampling/machine learning (ML) strategy.
• Utilized ML to create an accurate classical reference potential.
• Employed ML potential to guide long, collective Monte Carlo moves (nested Monte Carlo).
• Required minimal upfront computation for ML model training.

Main Results:

• Achieved AIMC performance levels matching or surpassing AIMD.
• Demonstrated high performance and exact sampling even with imperfect ML potentials.
• Generated statistically uncorrelated atomic configurations using very few ab initio calculations.

Conclusions:

• The hybrid nested sampling/ML approach significantly accelerates AIMC simulations.
• This method offers a powerful alternative for efficient and accurate atomic-scale materials modeling.
• It enables precise simulations with reduced computational cost.