| Abstract Scope |
The microstructure controls the mechanical properties of steel, which in turn is dictated by the process parameters. Microstructural analysis, however, is an art that requires an expert’s eye to suggest process modification. In this talk, we will present a robust methodology of microstructural quantification that segments pearlite colonies in the microstructure of hypoeutectoid steel, their orientation, and the interlaminar spacing. The method employs a Convolutional Neural Network (CNN) based UNET architecture as a semantic classifier. To address the challenge of limited labeled training data, a novel approach generates synthetic microstructures by combining polycrystalline templates and cropped images of pearlite and ferrite. This synthetic dataset effectively trains the UNET model, ensuring high accuracy and robustness. This technique, validated through stereological methods, proves versatile and can be generalized to other lamellar structures, offering a universal training methodology for multiphase microstructure quantification based on supervised learning. |