Discovery of Spin-Crossover Materials with Equivariant Graph Neural Networks and Relevance-Based Classification.

Area of Science:

• Materials Science
• Computational Chemistry
• Machine Learning

Background:

• Spin-crossover (SCO) materials are crucial for next-generation electronic and quantum devices.
• Efficiently identifying novel SCO materials is a significant challenge in materials discovery.
• Current high-throughput screening methods have limitations in speed and accuracy.

Purpose of the Study:

• To develop an accelerated method for discovering spin-crossover materials.
• To create a specialized database and train a predictive machine learning model.
• To enhance the identification of potential SCO systems for device applications.

Main Methods:

• Screened the Cambridge Structural Database to compile a specialized dataset of 1439 materials.
• Computed spin-switching energies using density functional theory (DFT).
• Trained an equivariant graph convolution neural network (GCNN) to predict spin-conversion energy, achieving a test mean absolute error of 360 meV.
• Integrated a relevance-based classifier for candidate identification.

Main Results:

• Developed a specialized database of 1439 materials with computed spin-switching energies.
• Achieved a predictive accuracy for spin-conversion energy with a test mean absolute error of 360 meV.
• Demonstrated a nearly 4-fold improvement in identifying potential spin-crossover systems compared to conventional high-throughput screening.

Conclusions:

• The developed machine learning approach significantly accelerates the discovery of spin-crossover materials.
• This method offers a more efficient and accurate alternative to traditional screening for identifying materials for electronic and quantum devices.
• The integrated system provides a powerful tool for advancing materials science research and development.