Evaluating creep rupture life in austenitic and martensitic steels with soft-constrained machine learning

Abstract

Machine learning is extensively utilized for predicting creep rupture of high-temperature steels. Recently, five soft-constrained machine learning algorithms (SCMLAs) have been developed to enhance the extrapolation capabilities of machine learning. These SCMLAs were applied to the austenitic steel Sanicro 25, showing their potential. To improve SCMLAs, this study has introduced new guidelines that address temperature culling within the input range and temperature extrapolation beyond the input range. Leveraging these guidelines, the SCMLAs were extended to various austenitic and martensitic stainless steels. The predicted results of TP316H, the data of which is representative of austenitic stainless steels, were validated through error estimates. Furthermore, notable agreement has been reached for temperature culling and temperature extrapolation, as demonstrated for TP91 and TP92 martensitic steels. The effects of single casts and the temperature dependence of the predictions have been analyzed for the studied materials. Consistent results can be readily achieved through systematic evaluations of SCMLAs for extrapolating up to 300,000 h or three times the maximum experimental rupture time for the studied materials. It is demonstrated that SCMLAs can provide reliable creep rupture life prediction across various high-temperature materials.

Keywords

• Creep rupture extrapolation
• Soft-constrained machine learning
• Soft-constrained neural network
• Stainless steels
• Error estimate