About this Abstract |
Meeting |
2024 TMS Annual Meeting & Exhibition
|
Symposium
|
Fatigue in Materials: Fundamentals, Multiscale Characterizations and Computational Modeling
|
Presentation Title |
Fatigue Damage Prediction Using Graph Neural Networks on Microstructure Representations |
Author(s) |
Ali Riza Durmaz, Akhil Thomas, Christoph Eberl, Peter Gumbsch |
On-Site Speaker (Planned) |
Ali Riza Durmaz |
Abstract Scope |
Crack initiation governs high-cycle fatigue (HCF) life and makes HCF very susceptible to microstructural details. Predicting microscopic cyclic plasticity in polycrystals requires thorough microstructure representations. In this work, we compare phenomenological crystal plasticity models with graph machine learning models on the task of predicting local cyclic plasticity in a ferritic steel (EN1.4003). A workflow is presented which consists of a combined experimental and data post-processing pipeline to establish fatigue damage data sets efficiently. It entails fatigue testing of mesoscale specimens to increase damage detection sensitivity, data fusion through multi-modal registration, and deep learning damage localization. The resulting data fuels both crystal plasticity and graph machine learning efforts. For the latter, the pixel data is transcribed into graph representations of the microstructure which facilitates the training of graph neural networks. Interpretability techniques are applied to learn about driving forces for the formation of extrusions and protrusions in individual grains. |
Proceedings Inclusion? |
Planned: |