Performance prediction of sintered NdFeB magnet using multi-head attention regression models

Area of Science:

• Materials Science
• Data Science
• Chemical Engineering

Background:

• Sintered Neodymium-Iron-Boron (NdFeB) magnet preparation is complex, time-consuming, and expensive.
• Traditional machine learning models face challenges with large, high-dimensional datasets and hyperparameter tuning.
• Neural networks offer powerful nonlinear modeling but lack interpretability.

Purpose of the Study:

• To develop a more efficient and interpretable machine learning model for NdFeB magnet synthesis.
• To enhance the understanding of structure-property relationships in materials.
• To accelerate the optimization of material performance.

Main Methods:

• Collected 1,200 high-quality experimental data points for NdFeB magnets.
• Developed a multi-head attention regression model integrating an attention mechanism into a neural network.
• Utilized parallel data processing for accelerated training and inference.

Main Results:

• Achieved coefficients of determination of 0.97 for remanence and 0.84 for coercivity.
• Demonstrated reduced reliance on feature engineering and hyperparameter tuning.
• Showcased enhanced interpretability of the neural network model.

Conclusions:

• The multi-head attention regression model significantly improves the efficiency and accuracy of NdFeB magnet property prediction.
• This approach offers valuable insights into machine learning-based material modeling.
• The study paves the way for advanced multimodal NdFeB magnet models.