Low carbon steel CCT diagram prediction using machine learning

This paper presents an approach to predicting Continuous Cooling Transformation (CCT) diagrams of low-carbon steels using a mathematical model based on regression and classification. The method of digitizing CCT diagrams and its application taken from atlases and current scientific articles are given. The digitization method is based on reading the color value from the CCT diagram image and then converting the coordinates of the color position in accordance with the scale of the CCT diagram axes. It was assumed that CCT diagram consists of zones defined by the beginning and end of ferrite transformation, the beginning of pearlite transformation, the end of bainite transformation, and the beginning and end of martensite transformation. When developing a predictive mathematical model, an optimization algorithm was used to find a model with the best hyperparameters among classical machine learning models (k-Nearest Neighbors, Support Vector Machine, Linear/Logistic Regression) and based on decision trees (LightGBM, CatBoost). The model solves regression (temperature prediction) and classification (binary mask prediction) problems. The superimposition of a binary mask on the temperature vector made it possible to constrain the resulting phase transformation curve along the time axis. To build test CCT diagrams, a number of dilatometric studies of four steel grades were carried out. The new predictive approach made it possible to achieve satisfactory values of metrics on test CCT diagrams. The average absolute error did not exceed 20°C; the coefficient of determination was in the range of 0.55–0.86, but for the martensite transformation it took negative values, which can be explained by the initial approximation of the transformation by a polygonal chain; ROC AUC metric was at least 0.80. The research was conducted within the framework of the Russian Federation’s strategic academic leadership program «Priority-2030,» aimed at supporting the development programs of higher education institutions. The scientific project PRIOR/SN/NU/22/SP5/26, titled «Creating Innovative Digital Tools for Applying Applied Artificial Intelligence and Advanced Statistical Analysis of Big Data in Technological Processes of Metallurgical Production», was also part of this initiative.