Adaptive-precision potentials for large-scale atomistic simulations.

Area of Science:

• Computational Materials Science
• Atomistic Simulations
• Machine Learning in Physics

Background:

• Large-scale atomistic simulations require efficient interatomic potentials for energies and forces.
• Machine-learning (ML) potentials offer high precision, while traditional potentials provide speed for larger systems.
• A gap exists between precision and computational cost in current simulation methods.

Purpose of the Study:

• To develop a novel multi-resolution method combining traditional and ML potentials.
• To create an adaptive-precision potential for optimizing performance and accuracy in complex atomistic systems.
• To enable efficient large-scale simulations by dynamically adjusting computational precision.

Main Methods:

• Implemented a multi-resolution potential combining a classical force field (Embedded Atom Model) and an ML potential (Atomic Cluster Expansion).
• Developed an adaptive precision scheme based on local structure analysis, automatically updating per-atom precision during simulation.
• Integrated the method into the LAMMPS molecular dynamics simulator, including a load balancer for variable computational loads.
• Demonstrated the approach using copper as a model system.

Main Results:

• The adaptive-precision potential achieved high accuracy, representing ML potential forces with 10 meV/Å precision and energies exactly for precisely calculated atoms.
• Simulations of nanoindentation on 4x10^6 copper atoms over 100 ps showed a significant speedup.
• Achieved an 11.3 times speedup compared to a full ML potential simulation.

Conclusions:

• The developed multi-resolution, adaptive-precision potential effectively balances computational cost and accuracy for large-scale atomistic simulations.
• This method offers a pathway to simulate more complex systems with unprecedented detail and efficiency.
• The approach is generalizable to other combinations of traditional and ML potentials.