Application of soft constrained machine learning algorithms for creep rupture prediction of an austenitic heat resistant steel Sanicro 25

Abstract

Creep rupture extrapolation is crucial for high-temperature materials served in power plants. Many analytical models can be used for creep rupture analysis, and fundamental models are also available. Machine learning is also an alternative. However, unphysical prediction curves occur readily in common machine learning algorithms, where one must manipulate the best results or ignore the less satisfactory ones. Using just high regression coefficients and low errors is not enough to obtain high accuracy of the methods. Nevertheless, five soft constrained machine learning algorithms (SCMLAs), where soft constraints, stability analysis by culling long-time or low-stress data, extrapolation from short to long times, and errors of solutions and algorithms are considered, are used for creep rupture prediction in this work. The models can generate reasonable results for fitting all data, extrapolating from short to long times, and stability analysis for Sanicro 25 after a number of tests. The errors of solutions for all the analyses are in a quite reasonable range, including extrapolation and stability analysis. The average relative standard deviation of the five SCMLAs is less than 2.5% at three times the maximum experimental creep rupture time. Creep rupture strength of the austenitic stainless steel Sanicro 25 can be predicted quantitatively by taking the average predicted stresses of the five SCMLAs. The method can also be used for other high-temperature alloys with similar creep degradation mechanisms.