Structure-Property Linkage in Alloys Using Graph Neural Network and Explainable Artificial Intelligence.

Area of Science:

• Materials Science
• Computational Materials Science
• Artificial Intelligence in Materials

Background:

• Deep learning accelerates materials science by predicting microstructure-property relationships.
• Convolution neural networks (CNNs) can analyze 3D microstructures but demand extensive resources.
• Limitations exist in CNNs regarding network size and training time for complex material analyses.

Purpose of the Study:

• To train and evaluate a graph neural network (GNN) for predicting mechanical properties of Ni-Al alloys.
• To compare the efficiency and interpretability of GNNs against CNNs for microstructure analysis.
• To leverage explainable AI for deeper understanding of material property prediction.

Main Methods:

• Generated Ni-Al alloy microstructures using phase field modeling.
• Trained a graph neural network (GNN) on these microstructures to predict mechanical property evolution.
• Employed saliency analysis and Bayesian Inference for model interpretability and parameter determination.

Main Results:

• The GNN accurately predicted alloy strengthening across varying microstructure sizes and dimensions.
• GNN demonstrated superior performance compared to CNNs, requiring less GPU utilization.
• Explainable AI tools provided interpretable insights into the GNN's predictions.

Conclusions:

• GNNs offer an accurate, efficient, and interpretable method for extracting information from material microstructures.
• GNNs overcome limitations of CNNs regarding microstructure size and dimensional restrictions.
• This approach advances the prediction of material properties and enhances understanding through explainable AI.