Machine learning for predicting occurrence of interphase precipitation in HSLA steels

Abstract Nano-sized interphase precipitates, which form in ordered rows, are critical for HSLA (High Strength Low Alloy) steels, in achieving the desired strength needed for downgauging for light-weighting automotive structures. The occurrence of interphase precipitation (as opposed to random, heterogeneous precipitation inside grains) is dependent on a number of inter-connected parameters such as alloy chemistry, ferrite-growth rate, temperature, crystallography and precipitate shape and size. In this study, we employ data analysis based on the Machine Learning algorithms: Decision Tree and Random Forrest to predict when interphase precipitation occurs. The results show that Decision Tree could not provide an accurate prediction, even as the depth was increased, whereas Random Forrest was able to, after up-sampling the minority class, a score of 98% for Accuracy, Recall, Precision and F1, which is adequate for predicting the occurrences of interfacial precipitates. The importance of individual features on the artificial classification task was evaluated and the precipitate chemistry and morphology were recognised by algorithms as the most informative features for decision making. Because our final goal was to predict the interphase precipitation for samples outside our training set we used receiver operating characteristic curve (ROC) and we found a strong classification prediction as the area under the curve was 0.98. In addition, particles containing Mo with disk shape were predicted to be very likely to form interphase precipitates. The alloy composition range for Ti, Mo and Mn was derived through Gaussian kernel estimate distribution and 0.18 Ti, 0.44 Mo and 1.5 Mn (all in wt.%) had the highest peaks.