Machine Learning Models for Creep Rupture Prediction of Inconel 617

Abstract

Creep rupture data are not always available at the desired temperature or stress levels, and performing creep tests can be both time-consuming and expensive. Creep rupture data from various sources are often combined for modeling. However, such combined data may overlap or exhibit a wide scatter band because of different metadata factors. A small change in chemical composition may affect the creep properties, creating a large variation in the rupture data. Machine learning (ML) offers a way to model these variations by including metadata such as chemical compositions. This study applies a python-based machine learning approach to predict the creep rupture in the form of Larson–Miller parameters (LMPs) of Inconel 617. Data from eight different sources (General Electric Company (GE), Oak Ridge National Laboratory (ORNL), German High-Temperature Gas Cooled Reactor (HTGR), Huntington Alloy, Korea Atomic Energy Research Institute (KAERI), Argonne National Lab (ANL), Atomic Energy Commission, and advanced ultrasupercritical (A-USC) boiler material data) which encompass multiple heats are used. Pearson correlation coefficient (PCC) and Spearman correlation coefficient (SCC) are employed to rank the input features based on their correlation with the Larson–Miller parameter, followed by an assessment of feature selection. Seven different regression methods (random forest (RF) regression, linear regression (LR), K-nearest neighbor (KNN), least absolute shrinkage and selection operator (LASSO), support vector regression (SVR), gradient boosting (GB) regression, and extreme gradient boosting (XGB)) are used for model training. The data are randomly split into training and testing datasets where the resulting prediction model is validated against testing data that is not used in calibration. Fivefold cross-validation and model learning curves are analyzed to rank model performances. A two-stage hyperparameter tuning is performed on suitable models for improved accuracy and stability and to minimize overfitting risk.