A generalised deep learning-based surrogate model for homogenisation utilising material property encoding and...

Area of Science:

• Materials Science
• Computational Mechanics
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are increasingly used for microstructure analysis and property prediction.
• Existing CNN models struggle to incorporate crucial material property information.
• This limitation hinders accurate structure-property relationship learning.

Purpose of the Study:

• To develop a method for encoding material properties directly into microstructure images for CNN analysis.
• To enhance CNNs for predicting properties of fiber-reinforced composite materials.
• To improve the accuracy and physical admissibility of surrogate models.

Main Methods:

• Developed a novel technique to embed material properties into microstructure images.
• Implemented a CNN model trained on these enhanced images for fiber-reinforced composites.
• Utilized learning convergence curves and mean absolute percentage error for optimization.
• Enforced Hashin-Shtrikman bounds to ensure physically admissible predictions.

Main Results:

• The developed CNN model effectively learns the structure-property relationship, including material information.
• Optimal training sample sizes were determined using learning convergence.
• The model demonstrated generality by predicting properties for unseen microstructures and extrapolated domains.
• Enforcing Hashin-Shtrikman bounds significantly improved performance in extrapolated regions.

Conclusions:

• The proposed method successfully integrates material properties into CNN-based microstructure analysis.
• This approach enhances the predictive capability and reliability of surrogate models for composites.
• The findings offer a pathway to more accurate and physically grounded material property predictions using AI.