Structure to Property: Chemical Element Embeddings for Predicting Electronic Properties of Crystals

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Predicting crystal properties is crucial for materials discovery.
• Existing methods may have limitations in accuracy or computational cost.

Purpose of the Study:

• To develop a general-purpose machine learning model for predicting diverse crystal properties.
• To achieve high accuracy in predicting electronic properties such as band gap and Fermi level energy.

Main Methods:

• The model utilizes atomic representations to capture detailed information about atoms and their environments.
• It is designed for broad applicability beyond electronic properties.

Main Results:

• The model accurately predicts crystal properties, including Fermi level energy and band gap.
• Achieved accuracy for band gaps surpasses previously published results.
• Demonstrated low computational requirements for efficient high-throughput screening.

Conclusions:

• The developed machine learning model offers a powerful and efficient tool for materials research.
• Its flexible architecture facilitates implementation and interpretation for computational chemists.
• The model can be extended to predict a wider range of chemical descriptors.