Predicting stress, strain and deformation fields in materials and structures with graph neural networks

Area of Science:

• Computational materials science
• Artificial intelligence in engineering
• Machine learning for physics-based simulations

Background:

• Developing accurate and fast computational tools for complex physical phenomena is a persistent challenge.
• Machine learning (ML) is transforming simulations, moving towards an AI-based paradigm.
• Efficiently predicting complex material and structural behavior using AI remains an area of active research.

Purpose of the Study:

• To present a general AI-based framework using graph neural networks (GNNs) for simulating complex mechanical behavior in materials.
• To demonstrate the framework's ability to learn material responses from a limited dataset.
• To establish a flexible approach for connecting material microstructure and properties to physical responses.

Main Methods:

• Implementation of a deep learning model utilizing graph neural networks (GNNs).
• Harnessing the mesh-to-graph mapping for data representation.
• Training the model on a few hundred data points to learn material behavior.

Main Results:

• The GNN framework accurately predicts deformation, stress, and strain fields in diverse material systems like composites and metamaterials.
• The model successfully captures complex nonlinear phenomena, including plasticity and buckling instability.
• The AI model appears to learn underlying physical relationships between predicted fields.

Conclusions:

• The developed graph-based AI framework offers a flexible and efficient approach for surrogate modeling of material behavior.
• This method connects microstructural features, material properties, and boundary conditions to macroscopic physical responses.
• The research opens new possibilities for AI-driven simulations in materials science and engineering.