Machine Learning at the Atomic Scale.

Area of Science:

• Computational chemistry
• Materials science
• Statistical learning

Background:

• Machine learning (ML) is increasingly used in atomic-scale modeling for predicting energies, forces, and properties.
• Developing robust and efficient methods for representing atomic configurations is crucial for ML applications in this field.

Purpose of the Study:

• To review recent advancements in machine learning for atomic-scale modeling.
• To highlight methods for representing atomic configurations and regression algorithms for property prediction.
• To demonstrate how model optimization can uncover physical insights into structure-property relationships.

Main Methods:

• Focus on mathematically robust and computationally efficient atomic configuration representations.
• Discussion of various regression algorithms used for building surrogate models of atomic-scale properties.
• Illustrative examples of optimizing ML models to understand physical phenomena.

Main Results:

• Recent progress in machine learning for predicting molecular and condensed-phase properties.
• Effective strategies for representing atomic structures in a computationally feasible manner.
• Demonstration of ML model optimization revealing underlying physical principles.

Conclusions:

• Machine learning is a powerful tool for atomic-scale modeling, enhancing prediction accuracy and efficiency.
• The representation of atomic configurations is key to successful ML model development.
• Optimizing ML models provides valuable insights into the physics governing material properties.