Machine Learning Prediction of Cyclic Oxidation Behavior in Superalloys: A New Breakthrough in Materials Informatics

Based on ensemble regression models (e.g., CatBoost) and SHAP interpretability analysis, a framework for predicting mass change during cyclic oxidation of superalloys is constructed. This addresses the issues of high time and cost associated with traditional empirical formulas and laboratory tests, providing an intelligent analysis tool for materials science.

In fields such as aerospace and energy power generation, cyclic oxidation of superalloys (involving oxide scale spallation caused by thermal stress) directly affects component lifespan and safety, while traditional evaluation methods are inefficient. The rise of materials informatics provides new ideas for machine learning to solve this problem.

cyclic-oxidation-ml framework: Predicts cyclic oxidation behavior of Fe–Cr binary and Fe–Cr–Ni ternary alloys; Dataset: Includes alloy composition (Fe/Cr/Ni percentages), environmental parameters (temperature/exposure time), and target variable (mass change); Model comparison: Evaluated 12 regression algorithms, with ensemble methods (CatBoost/Extra Trees/Random Forest) outperforming linear models.

• Temperature dominance: Increasing temperature accelerates oxidation kinetics and oxide scale spallation;
• Chromium protection: Forms a dense Cr₂O₃ film; its content is positively correlated with oxidation resistance but has a saturation threshold;
• Nickel synergy: Improves oxide scale adhesion and works synergistically with chromium to enhance performance;
• Time-temperature interaction: Rapid mass gain occurs at high temperatures for short durations, while mass loss may occur at medium temperatures over long durations due to spallation.

It was found that duplicate samples plus random splitting can inflate R² (close to 0.99), revealing that ML applications in materials science need to pay attention to:

• Data leakage risks (correlated data caused by similar experimental conditions);
• The necessity of strict cross-validation;
• The final cross-validated CatBoost model has an R² of approximately 0.88, showing good generalization performance.

Potential applications: New material screening, process optimization, component lifespan prediction; Future directions: Integrate CALPHAD thermodynamic calculations, DFT electronic structure descriptors, and physics-informed neural networks, and expand to nickel-based/cobalt-based alloy systems.

This project demonstrates the potential of machine learning in materials science. Through rigorous validation and interpretability analysis, it not only achieves high-precision prediction but also reveals the mechanism of action of key factors, providing a feasible path for new material development.