Process of building machine learning models to predict microstructures of pipe steels after continuous cooling involves the collection and preparation of data, the source of which is thermokinetic diagrams of supercooled austenite decomposition. Statistics of intermediate and final data, as well as algorithms for their transformation are given. Evaluations of machine learning models for selected microstructures are considered. A method for generating data under small sample conditions and introducing an evaluative feature of grain size are proposed. Models were validated and the significance of features was interpreted. The practical use of models for constructing thermokinetic diagrams of austenite decomposition and analysis of modeling results is shown.