Machine-learning investigation on the creep-rupture time of heat-resistant steels

Abstract

Currently, the reliable prediction of creep-rupture times of heat-resistant alloys is a major research area with regard to structural materials. Numerous efforts to predict creep rupture times have been made using phenomenological parameter methods; however, these approaches have only been able to achieve a limited prediction accuracy. Furthermore, the applicability of machine learning to improve the prediction accuracy of creep-rupture time has been investigated. In both studies, nevertheless, approaches have been used where explicit analytical formulas are not available. Because creep rupture is a safety issue, it is essential to use an analytical equation for predicting rupture time. Therefore, in this study, a machine learning method was applied to the analysis of creep-rupture time of heat-resistant steel to obtain an explainable prediction formula. An accurate regression model was obtained using the linear independent descriptor generation (LIDG) regression method of adding independent higher-order terms that were generated from operations between variables. By utilizing LIDG regression, we show that various creep-rupture times could be well-reproduced by different heats using explanatory variables such as creep conditions, chemical compositions, and Vickers hardness. The heat-to-heat variation in creep-rupture time could not be accounted for at all in the evaluation using the conventional parameter method. On the other hand, the LIDG regression could naturally explain the difference in the creep rupture time of various heats. Furthermore, the excellent extrapolation of the LIDG regression model was confirmed, showing its potential for long-time creep-rupture time prediction updating the conventional parametric method.