Explainable machine learning for materials discovery: predicting the potentially formable Nd-Fe-B crystal structures...

Area of Science:

• Materials Science
• Computational Materials Science
• Crystallography

Background:

• Neodymium-Iron-Boron (Nd-Fe-B) magnetic materials are crucial for permanent magnets.
• Exploring novel Nd-Fe-B crystal structures is essential for enhancing magnetic properties.
• Elemental substitution in lanthanide-transition metal-light element (LA-T-X) systems offers a pathway to new materials.

Purpose of the Study:

• To computationally design and evaluate novel Nd-Fe-B crystal structures.
• To develop and compare data-driven models for predicting phase stability.
• To identify key structural descriptors influencing the stability of new Nd-Fe-B phases.

Main Methods:

• High-throughput first-principles calculations were used to assess phase stability.
• A dataset of 5967 ternary LA-T-X host structures was generated.
• Supervised (kernel ridge regression, logistic classification, decision tree) and unsupervised (descriptor-relevance analysis with Gaussian mixture model) learning techniques were applied for stability prediction.

Main Results:

• 20 potentially formable Nd-Fe-B crystal structures were identified through first-principles calculations.
• Supervised learning models achieved maximum accuracy of 70.4% and recall of 68.7%.
• The proposed unsupervised learning model significantly outperformed supervised models, achieving 72.9% accuracy and 82.1% recall.

Conclusions:

• Unsupervised learning provides a superior approach for predicting the stability of novel Nd-Fe-B crystal structures.
• Average atomic coordination number and Fe site coordination number are critical factors for phase stability.
• This study paves the way for accelerated discovery of advanced magnetic materials.