Designing complex architectured materials with generative adversarial networks.
Yunwei Mao1, Qi He1, Xuanhe Zhao1,2
1Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA 02139, USA.
Science Advances
|June 5, 2020
Summary
This study introduces an AI-driven method for designing architectured materials, achieving extreme elastic properties without prior design knowledge. This approach enables the creation of advanced materials for diverse applications.
Area of Science:
- Materials Science
- Mechanical Engineering
- Artificial Intelligence
Background:
- Architectured materials offer desirable properties for various applications.
- Current design methods rely on expert knowledge, limiting broader use.
- Designing materials with extreme properties, like maximal elastic stiffness, is challenging.
Purpose of the Study:
- To develop an experience-free method for designing complex architectured materials.
- To achieve architectured materials approaching theoretical elastic property bounds.
- To overcome limitations of traditional design approaches.
Main Methods:
- Utilized generative adversarial networks (GANs) for an experience-free design approach.
- Trained GANs on simulation data from millions of randomly generated architectures.
- Categorized architectures based on crystallographic symmetries for training.
Main Results:
- Successfully designed over 400 two-dimensional architectured materials.
- Achieved material properties approaching the Hashin-Shtrikman upper bounds for isotropic elastic stiffness.
- Demonstrated effectiveness across a wide range of porosities (0.05 to 0.75).
Conclusions:
- The proposed GAN-based approach enables systematic and experience-free design of architectured materials.
- This method can create materials with extreme mechanical properties, pushing performance boundaries.
- The findings pave the way for mass production of advanced architectured materials via additive manufacturing.
