Machine-learned potentials for next-generation matter simulations.

Area of Science:

• Computational Materials Science
• Machine Learning Applications

Background:

• The fundamental challenge in computational materials science is balancing simulation accuracy with computational cost for large-scale phenomena.
• Accurate simulations are crucial for investigating complex materials behaviors across significant time and length scales.

Purpose of the Study:

• To review the emerging field of machine-learned potentials (ML potentials) in computational materials science.
• To summarize the principles, data acquisition, and active learning procedures of ML potentials.
• To highlight diverse applications and discuss future developments for broader adoption.

Main Methods:

• Review of machine learning principles applied to interatomic potentials.
• Discussion of data acquisition strategies for training ML potentials.
• Summary of active learning procedures for efficient model refinement.

Main Results:

• Machine-learned potentials achieve quantum mechanical accuracy at a reduced computational cost.
• ML potentials have demonstrated success in diverse fields including organic chemistry, biomolecules, crystal structure prediction, and surface science.
• The review highlights the potential of ML potentials for advancing materials design and solving key scientific challenges.

Conclusions:

• Machine-learned potentials represent a significant advancement, bridging the gap between accuracy and computational efficiency.
• Further development is needed to promote wider adoption and unlock the full potential of ML potentials in materials science.
• ML potentials are poised to facilitate fully computational materials design and address fundamental questions in the field.