Apr 10, 2025
4:00pm - 4:15pm
Summit, Level 4, Room 422
Min Jik Kim1,2,Woo Seok Yang1,Sang Min Park2,Da Seul Shin1
Korea Institute of Materials Science1,Pusan National University2
Min Jik Kim1,2,Woo Seok Yang1,Sang Min Park2,Da Seul Shin1
Korea Institute of Materials Science1,Pusan National University2
It is essential to understand the high-temperature deformation behavior under various processing conditions for manufacturing components from hard-to-deform materials. This behavior is highly nonlinear and complex with respect to deformation temperatures and strain rates. During hot deformation, the true stress-strain curves exhibit different trends due to work hardening and softening, accompanied by corresponding microstructural changes. According to previous studies, various deep learning models have been demonstrated to predict flow stresses; however, these models only predict discrete flow stress data points during the hot forming process and have not been applied to capture or utilize microstructural evolution.
In this study, deep learning approaches are proposed to describe (i) the flow behavior and (ii) the microstructural evolution related to the hot deformation of Inconel 718 alloy under various processing conditions, based on experimental flow curves obtained from uniaxial hot compression tests and crystallographic microstructures from deformed samples using electron backscatter diffraction (EBSD).
First, the autoencoder-prediction network (AE-PN) is a robust model for predicting flow stress curves under high-temperature plastic deformation. By learning the non-linearly reduced latent vectors from flow curves in conjunction with the processing parameters, it enables the prediction of a continuous flow curve. In addition, it allows for the analysis of the latent space as interpretable, physics-driven data for each curve.
Second, the conditional variational autoencoder (CVAE) was developed to generate the latent space for microstructural evolution based on microstructure images and thermo-mechanical processing conditions. By exploring the latent space, virtual microstructures can be randomly generated, and the microstructural analysis networks were used to examine the synthetic images to predict kernel average misorientation (KAM) distribution and to detect deformed grains. The CVAE can identify the cluster with the target synthetic images and the corresponding thermo-mechanical processing conditions through a genetic algorithm (GA) optimization-based exploration of the latent space. Furthermore, the identified conditions could be located in regions with high power dissipation values within the processing map.
In summary, autoencoder-based deep learning models provide new insights as hot deformation behavior descriptors by not only predicting flow behavior under various forming conditions but also generating deformed microstructures.