Uncertainty quantification for Bayesian active learning in rupture life prediction of ferritic steels

Area of Science:

• Materials Science
• Mechanical Engineering
• Computational Science

Background:

• Creep rupture life prediction is critical for high-temperature applications of 9-12%Cr ferritic-martensitic steels.
• Accurate modeling requires accounting for material composition and processing parameters.
• Quantifying prediction uncertainty is essential for risk assessment.

Purpose of the Study:

• Develop and compare probabilistic methodologies for predicting long-term creep rupture life.
• Evaluate the efficacy of machine learning models in capturing creep behavior and associated uncertainties.
• Investigate the potential of active learning for efficient material exploration.

Main Methods:

• Developed three probabilistic methodologies for creep life prediction.
• Applied Gaussian Process Regression (GPR), quantile regression, and natural gradient boosting.
• Evaluated model performance using holdout test sets and quantified epistemic uncertainty.
• Simulated an active learning framework for experimental data collection.

Main Results:

• Gaussian Process Regression demonstrated superior inference accuracy (R² value) and meaningful uncertainty estimation (94-98% coverage).
• GPR outperformed quantile regression and natural gradient boosting algorithms in predicting creep rupture life.
• The active learning framework simulation showed potential for intelligent material space exploration.

Conclusions:

• Probabilistic machine learning, particularly GPR, provides an effective framework for predicting creep rupture life and its uncertainty in 9-12%Cr steels.
• Addressing epistemic uncertainty is crucial for developing accurate and reliable predictive models.
• Active learning strategies can significantly reduce experimental effort in material development and optimization.