Extending machine learning beyond interatomic potentials for predicting molecular properties.

Area of Science:

• Computational Chemistry
• Materials Science
• Artificial Intelligence

Background:

• Machine learning (ML) is emerging as a powerful tool for modeling complex chemical systems.
• ML models act as surrogate models, learning relationships between molecular structure and chemical properties from data.

Purpose of the Study:

• To review recent advancements in applying ML to chemical property evaluation.
• To explore ML's role in obtaining reduced quantum-mechanical descriptions.
• To highlight the development of ML architectures and their applicability in chemistry.

Main Methods:

• Overview of modern neural network architectures for chemical property prediction.
• Evaluation of ML model predictive capabilities, generality, and transferability.
• Discussion of ML's ability to represent molecular structures and capture quantum mechanical effects.

Main Results:

• Learned molecular representations in ML models approximate quantum-mechanical analogues.
• ML models demonstrate capability in describing non-local quantum effects.
• Applicability of ML to diverse chemical properties like partial atomic charges and chemical bonding.

Conclusions:

• The field is progressing towards physics-informed ML models for chemistry.
• Development of new methods and user-friendly ML frameworks is accelerating.
• Future outlook includes addressing challenges and expanding ML applications in chemical sciences.