Accelerating atomistic simulations with piecewise machine-learned ab Initio potentials at a classical force...

Area of Science:

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

Background:

• Machine learning (ML) enables high-dimensional interatomic potentials with ab initio accuracy.
• Current ML potentials are faster than first-principles calculations but slower than classical force fields due to complex descriptors.

Purpose of the Study:

• To bridge the efficiency gap between ML potentials and classical force fields.
• To develop a computationally efficient ML interatomic potential suitable for large-scale simulations.

Main Methods:

• Proposed an embedded atom neural network approach.
• Utilized simple piecewise switching function-based descriptors for linear scaling with neighbor atoms.
• Validated the model on metallic and covalent materials.

Main Results:

• Achieved over an order of magnitude speedup compared to popular ML potentials with comparable accuracy.
• The model's performance approaches the speed of the fastest embedded atom method.
• Demonstrated favorable linear scaling with the number of neighbor atoms.

Conclusions:

• The proposed piecewise ML model offers extreme efficiency for atomistic simulations.
• This approach is promising for simulating very large systems and/or long timescales in materials science.
• Enables faster and accurate materials modeling through advanced computational techniques.