Creep Life Predictions by Machine Learning Methods for Ferritic Heat Resistant Steels

We have attempted to predict creep rupture time for a wide range of ferritic heat resistant steels with machine learning methods using the NIMS Creep Data Sheets (CDSs). The datasets consisted of commercial-steel data from 27 CDSs, including those on various grades of carbon, low- alloy, and high-Cr steels. The prediction models were constructed using three methods, namely, support vector regression (SVR), random forest, and gradient tree boosting with 5132 training data, to predict log rupture time from the chemical composition (19 elements), test temperature, and stress. Evaluation with 451 test data proved that all three models exhibited a high predictivity of creep rupture time. In particular, the performance of the SVR model was the highest with a root mean squared error as low as 0.14 over the log rupture time; this value means that, on average, the prediction error had a factor of 1.38 (=100.14). The high predictivity achieved without using microstructure information was presumably due to the fact that the data used were for commercial steels in which there should be a correlation between the chemical composition and the microstructure. We confirmed that the prediction did not work exceptionally well for two heats having the same composition but different microstructures with and without stress-relief annealing. The predictivity could be markedly increased by adding the 0.2% proof stress at the creep test temperature as one of the explanatory variables. As a demonstration of the prediction model, the effect of elements was evaluated in modified 9Cr–1Mo steels.