Characterization and Inverse Design of Stochastic Mechanical Metamaterials Using Neural Operators.

Area of Science:

• Materials Science
• Mechanical Engineering
• Artificial Intelligence

Background:

• Machine learning (ML) offers advanced capabilities for designing mechanical metamaterials.
• Current inverse design methods struggle with data-intensive requirements, especially for nonlinear micro-architected materials.
• Designing stochastic architected materials with nonlinear mechanical behaviors is challenging due to data limitations.

Purpose of the Study:

• To develop a comprehensive ML framework for inverse design of mechanical metamaterials using sparse experimental data.
• To leverage deep neural operators (e.g., DeepONet) for learning microstructure-property relationships.
• To enable efficient design of materials with specific nonlinear mechanical responses.

Main Methods:

• A scientific ML framework employing deep neural operators (DeepONet and variants) was developed.
• The framework learns the relationship between microstructure and mechanical response from sparse, high-quality in situ experimental data.
• Systematic comparison of various neural operators and standard neural networks was performed for interpretability and accuracy.

Main Results:

• The ML framework successfully learned the relationship between microstructure and mechanical response.
• Prediction errors for mechanical responses of stochastic spinodal microstructures were within 5-10%.
• The approach demonstrated efficient inverse design capabilities for targeted nonlinear mechanical behaviors.

Conclusions:

• Deep neural operators, combined with advanced mechanical experiments, facilitate the design of complex micro-architected materials.
• The framework is effective even with data scarcity, enabling the design of metamaterials with desired nonlinear properties.
• This work advances materials-by-design, paving the way for next-generation metamaterials informed by experimental insights.