Materials fatigue prediction using graph neural networks on microstructure representations.

Area of Science:

• Materials Science
• Mechanical Engineering
• Computational Science

Background:

• Predicting fatigue damage in polycrystals under high-cycle fatigue is a persistent challenge.
• Identifying grains prone to plastic deformation under cyclic loading is crucial for understanding material failure.
• Existing methods struggle to accurately capture localized damage initiation within complex microstructures.

Purpose of the Study:

• To develop a novel approach for predicting local fatigue damage in polycrystals.
• To leverage graph neural networks (GNNs) for grain-wise damage classification in ferritic steel.
• To identify effective data representations and GNN models for understanding fatigue damage mechanisms.

Main Methods:

• Microtexture and damage maps from experimental data were transcribed into a microstructure graph representation.
• Grains were represented as nodes, with edges connecting adjacent grains.
• Graph convolutional networks (GCNs) were applied for binary grain-wise damage classification.

Main Results:

• Graph convolutional networks achieved a balanced accuracy of 0.72 and an F1-score of 0.34.
• GCNs significantly outperformed phenomenological crystal plasticity (+68%) and conventional machine learning (+17%) models.
• Interpretability analysis highlighted critical grains and features influencing fatigue damage initiation.

Conclusions:

• GNNs offer a powerful tool for predicting fatigue damage initiation in polycrystals.
• This approach can reveal underlying microstructural driving forces and mechanisms of fatigue failure.
• The microstructure graph framework facilitates the application of advanced machine learning techniques to materials science problems.