EXTENDED ABSTRACT: Designing materials to meet specific demands and identifying suitable manufacturing processes is a fundamental task in materials science. Achieving this goal requires a deep understanding of the relationship between process, structure, and property (PSP). Traditional methods involve feature engineering to identify significant features that impact subsequent steps. In contrast, deep learning approaches provide a framework free from feature engineering while delivering excellent performance and generalizability. This study introduces a deep learning framework to establish the PSP linkage for the heat treatment process, microstructures, and mechanical properties of 42CrMo4 steel. We used a conditional StyleGAN algorithm to generate microstructure images based on tempering temperatures, and a ResNet algorithm to predict yield strength, tensile strength, and elongation based on microstructural images. Samples of 42CrMo4 steel were heat-treated at various tempering temperatures. Microstructures were observed with an optical microscope, and mechanical properties were measured. Using these datasets, two deep learning models were applied. First, the conditional StyleGAN was trained with tempering temperatures and corresponding microstructures. Then, the ResNet model was trained with microstructural images as input to predict mechanical properties. At low tempering temperatures, the microstructure is mainly composed of tempered martensite, while the proportion of ferrite increases as the temperature rises. This trend was also observed in the generated images with the conditional StyleGAN. Strength values decreased with increasing tempering temperature for both observed and generated images. The strengths predicted for the generated microstructures with the ResNet model corresponded well with those for the observed microstructures. In addition, this framework is implemented with unknown tempering temperature conditions and comparing the results to those obtained under given experimental conditions, it was demonstrated that this framework can produce plausible microstructural images and accurately predict properties.
Keywords: PSP relationship; steels; deep learning; microstructure; mechanical properties
REFERENCES:
[1] K. Tero et al., Proc. Of IEEE CVPR, (2019), 4401.
[2] H. Kaiming et al., Proc. Of IEEE CVPR, (2016), 770
Hoheok Kim is a senior researcher at the Korea Institute of Materials Science (KIMS). He earned his B.S. in Materials Science and Engineering from Sunkyunkwan University and his Ph.D. in the same field from the University of Tokyo. He worked at LG Chem before joining KIMS, where he is currently part of the Materials Data & Analysis Research Division. His main research interests focus on the development and application of artificial intelligence for materials analysis and property prediction. His past research includes applying machine learning methods to phase transformation analysis, using deep learning for microstructure segmentation, and modeling the property-structure relationship.