Efficient and accurate spatial mixing of machine learned interatomic potentials for materials science

Area of Science:

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

Background:

• Machine-learned interatomic potentials (MLIPs) offer high accuracy but are computationally intensive.
• Existing methods struggle with large-scale simulations due to computational cost.
• Quantum mechanics/molecular mechanics (QM/MM) inspired approaches can bridge accuracy and efficiency gaps.

Purpose of the Study:

• To introduce ML-MIX, a novel package for accelerating molecular dynamics simulations.
• To enable the use of complex MLIPs on larger systems within practical computational budgets.
• To demonstrate the accuracy and efficiency gains of spatially mixing interatomic potentials.

Main Methods:

• Developed ML-MIX, a LAMMPS package compatible with CPUs and GPUs.
• Implemented a spatial mixing strategy for interatomic potentials of varying complexity.
• Demonstrated potential distillation techniques to create 'cheap' potentials from 'expensive' ones.
• Validated performance on point defects in Si, Fe, and W-He systems.

Main Results:

• Achieved speedups of up to 11× for systems around 8000 atoms without loss of accuracy.
• Successfully applied ML-MIX to study screw dislocation mobility in W using ACE potentials.
• Accurately reproduced experimental He reflection coefficients in W using MACE/SNAP mixing, matching observations up to 80 eV.
• Showcased the capability of deploying state-of-the-art MLIPs on large, realistic systems.

Conclusions:

• ML-MIX significantly accelerates atomistic simulations by efficiently combining MLIPs.
• The method allows for the application of advanced MLIPs to large-scale, scientifically relevant problems.
• ML-MIX bridges the gap between computational cost and accuracy in materials simulations, enabling new discoveries.