Iterative training set refinement enables reactive molecular dynamics via machine learned forces.

Area of Science:

• Computational Chemistry
• Materials Science
• Surface Physics

Background:

• Machine learning (ML) shows promise in computational chemistry and physics.
• Atomistic simulations using ML-driven forces remain a significant challenge.

Purpose of the Study:

• To demonstrate the efficacy of neural network-trained forces for simulating reactive self-sputtering from a beryllium surface.
• To develop and implement a refinement protocol for improving the extrapolation capabilities of ML potentials.

Main Methods:

• Utilized neural network potentials trained on density functional theory (DFT) data.
• Implemented an iterative refinement protocol to enhance molecular dynamics (MD) simulations.
• Validated results against ab initio MD simulations and established potential models.

Main Results:

• Achieved accurate simulation of reactive self-sputtering from beryllium surfaces.
• Sputtering yield results for incident energies below 100 eV closely matched ab initio MD simulations.
• Demonstrated that ML potentials require augmented training data for different crystallographic surfaces (e.g., Be(0001) and Be(011̄0)).

Conclusions:

• ML-trained forces offer a viable and accurate alternative for atomistic simulations of surface processes like sputtering.
• The developed refinement protocol addresses limitations in ML potential extrapolation.
• This method allows for larger, longer, and more statistically relevant simulations than direct ab initio MD.