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 Sheet (CDS). The dataset consisted of commercial steel data from 27 sheets in the CDS, covering various grades of carbon steels, low alloy steels, and high Cr steels. The prediction models were constructed using three methods, support vector regression (SVR), random forest, and gradient tree boosting with 5132 training data in order to predict log rupture time from chemical composition (19 elements), test temperature, and stress. Evaluation with 451 test data proved that all three models exhibited 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, which value means that, on average, the prediction error was factor 1.38 (=100.14). The high predictivity achieved with no use of information on microstructure was presumably because the data used was for commercial steels in which there should be a correlation between the composition and the microstructure. We confirmed that the prediction did not work well exceptionally for two heats having the same composition but different microstructures with and without stress relief annealing. The predictivity could be drastically improved by adding the 0.2% proof stress at the creep test temperature as one of the explanatory variables. As a use case of the prediction model, the effect of elements was evaluated for modified 9Cr 1Mo steels.