Automated discovery of a robust interatomic potential for aluminum

Area of Science:

• Computational Materials Science
• Quantum Mechanics
• Machine Learning

Background:

• Machine learning (ML) potentials trained on quantum mechanics (QM) calculations are crucial for modeling potential energy surfaces.
• The accuracy of these ML potentials heavily relies on the quality and diversity of the training dataset.

Purpose of the Study:

• To present a highly automated approach for constructing training datasets for ML potentials.
• To demonstrate this method by developing an ML potential for elemental aluminum (ANI-Al).

Main Methods:

• An active learning scheme was employed, using the ML potential to drive non-equilibrium molecular dynamics simulations.
• New QM data was collected when ML uncertainty was high, and the ML model was periodically retrained.
• The method involves automated dataset construction and active learning for QM data selection.

Main Results:

• The developed ANI-Al potential accurately predicts the radial distribution function in melt and the liquid-solid coexistence curve.
• It also shows high accuracy for crystal properties, including defect energies and barriers.
• A 1.3M atom shock simulation demonstrated excellent agreement between ANI-Al force predictions and reference DFT calculations.

Conclusions:

• The automated dataset construction and active learning approach is effective for developing accurate ML potentials.
• The ANI-Al potential provides reliable predictions for aluminum properties across different phases and conditions.
• This method facilitates the creation of robust ML models for materials simulations.