Data-Driven Fatigue Prediction of Superalloys: A Novel Strategy Integrating Transfer Learning and Partial Label...

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning Applications

Background:

• Machine learning (ML) offers efficient material property prediction but struggles with ambiguous data.
• Data integrity issues, especially ambiguous data, hinder the widespread application of ML in materials science.
• Predicting fatigue performance in superalloys is critical for engineering applications.

Purpose of the Study:

• To develop a novel ML strategy for predicting superalloy fatigue performance using ambiguous compositional data.
• To enhance model interpretability by uncovering composition-microstructure-property relationships.
• To establish a robust framework for materials property prediction applicable to broader datasets.

Main Methods:

• Integration of partial label learning and transfer learning to manage ambiguous compositional data.
• Enrichment of microstructural features via thermodynamic calculations based on composition.
• Development of ML models for fatigue performance prediction in superalloys.

Main Results:

• Achieved superior predictive accuracy for superalloy fatigue performance.
• Demonstrated robust generalization capabilities validated through experimental data.
• Enhanced model interpretability by revealing underlying composition-microstructure-property correlations.

Conclusions:

• The proposed framework effectively addresses ambiguous data challenges in materials property prediction.
• The integration of partial label and transfer learning offers a powerful approach for materials informatics.
• This methodology has significant potential for advancing materials design and discovery.