A MACHINE LEARNING APPROACH FOR STRESS-RUPTURE PREDICTION OF HIGH TEMPERATURE AUSTENITIC STAINLESS STEELS

This study outlines a machine learning approach for long-term stress-rupture (SR) prediction of high temperature austenitic stainless steel. Traditional methods of lifetime estimation and alloy design for turbomachinery application rely mostly on repeated testing, prior experience, and trial-and-error approach, which are laborious, time intensive, and costly. Recent advances in machine learning offer an accelerated technique for the development of constitutive creep laws, superior alloy designs, and reliable long-term performance prediction. To that end, a machine learning approach is explored in this study for long-term stress-rupture prediction. The toolbox GPTIPS, a biologically inspired genetic programming (GP) algorithm for building accurate and intrinsically explainable non-linear regression model is employed in this study. In GPTIPS, randomly sampled tree structures, mutate and cross over the best performing trees to create a new sample. The process iterates until the best solution is found based on criteria set by the user. Herein, the stress-rupture data of 18Cr-8Ni (304 SS) stainless steel, divided into 60% training and 40% testing data irrespective of heat grades are feed into GPTIPS. The GPTIPS is iterated based on the number of genes, tournament size, tree depth, and nodes. The generated SR constitutive models are ranked according to goodness-of-fit and model complexity. The best-ranked models are compared with the experimental data and found to be free of inflection points at low-stress. Post audit validation is performed by fitting the model blindly against an extended data base of 18Cr-12Ni-Mo (316 SS) stainless steel. Based on the goodness-of-fit, the best-ranked models are investigated for future application, comprehensive understanding of their limitations, and the resultant capability of effective prediction. In future work, the ability of GPTIPS will be leveraged to develop minimum-creep-strain-rate models, alloy design based on chemical composition, potential sources of uncertainty, and their implications on the outcomes.