Perspective: Machine learning potentials for atomistic simulations.

Area of Science:

• Computational Chemistry
• Condensed Matter Physics
• Materials Science

Background:

• Computer simulations are essential tools in chemistry, physics, and materials science.
• Increasing demand for accurate simulations of larger, more complex systems.
• Development of efficient interatomic potentials is a critical bottleneck.

Purpose of the Study:

• To review the paradigm shift towards machine learning potentials (ML potentials).
• To discuss the underlying principles, successes, and challenges of ML potentials.
• To assess the current applicability and limitations of ML potentials in simulations.

Main Methods:

• Fitting large datasets from electronic structure calculations.
• Utilizing machine learning algorithms to represent potential-energy surfaces.
• Comparison with traditional, physics-based interatomic potentials.

Main Results:

• Machine learning provides an alternative to traditional potentials for accurate energy and force calculations.
• ML potentials can represent complex potential-energy surfaces effectively.
• This approach addresses limitations of simplified potentials used in large-scale simulations.

Conclusions:

• Machine learning potentials represent a significant advancement in computational modeling.
• Ongoing research is crucial to overcome remaining challenges and expand applicability.
• ML potentials are becoming vital for state-of-the-art simulations in various scientific fields.