A Hybrid Stacked Ensemble Learning Framework Using a Time-Temperature Parameter Metamodel for Improved Creep Rupture Prediction

Abstract Scarce creep-rupture data at the desired temperature and stress levels often necessitates combining datasets from multiple sources to achieve accurate predictions. However, such data aggregation may lead to overlapping or wide scatter bands due to varying metadata factors. The traditional modeling approaches are not suitable to incorporate metadata for rupture prediction and are unable to capture the complex relationship between these variables and rupture life. To address these challenges, this study introduces a hybrid method that integrates twelve Time-Temperature parameters (TTP) with stacked ensemble learning to improve creep rupture predictions. First, the experimental creep rupture data are transformed into twelve TTPs, producing a normal distribution suited for machine learning applications. Next, stacked ensemble learning is applied to these twelve TTP datasets, combining the strengths of multiple algorithms. Four ensemble models are used as base learners for model training. The outputs of these base learners are then combined using a meta-learner to identify the optimal model combination. The proposed approach is validated using an aggregated dataset of 469 creep-ruptures of alloy Inconel 617 which were collected from eight different sources. The results show that the hybrid strategy (1) avoids the bias of using a preselected TTP model, and (2) improves prediction performance against aggregated data from multiple sources compared to the traditional approaches.