Atomic Energies from a Convolutional Neural Network.

Area of Science:

• Computational Materials Science
• Machine Learning in Chemistry
• Quantum Mechanics

Background:

• Understanding atomic-level interactions is crucial for designing new materials.
• First-principles calculations offer detailed insights but are computationally expensive.
• Machine learning (ML) shows promise in predicting material properties efficiently.

Purpose of the Study:

• To introduce a new, parameter-free structure descriptor called 'k-Bags'.
• To present a scalable ML framework for analyzing atomic properties of structures.
• To enable accurate prediction of structure-energy relations and atomic energies.

Main Methods:

• Development of the 'k-Bags' simplified structure descriptor.
• Implementation of a comprehensive and scalable machine learning framework.
• Integration of chemically meaningful atomic energies into global structure optimization.

Main Results:

• The ML model accurately predicts structure-energy relations, comparable to ab initio methods.
• Chemically meaningful atomic energies are generated for organic and inorganic structures.
• Global structure optimization is significantly accelerated using local atomic energy information.

Conclusions:

• The 'k-Bags' framework offers a computationally efficient alternative to traditional methods.
• This approach deepens the understanding of atomic properties in diverse molecular and material structures.
• The method accelerates the search for global minimum structures in complex systems.