Inverting the structure-property map of truss metamaterials by deep learning.

Area of Science:

• Materials Science
• Mechanical Engineering
• Computational Science

Background:

• Architected materials based on truss lattices offer beneficial mechanical properties.
• Inverse design of these materials, particularly for tailored stiffness, remains a significant challenge.
• Current methods struggle to efficiently identify structures with specific homogeneous or spatially varying properties.

Purpose of the Study:

• To develop a deep-learning framework for predicting truss architectures with fully tailored anisotropic stiffness.
• To overcome the limitations of current inverse design methods in architected materials.
• To enable the creation of materials with precisely controlled mechanical responses.

Main Methods:

• A deep-learning framework combining neural networks with enforced physical constraints was developed.
• The framework was trained on millions of unit cells to cover a vast design space.
• The model accurately predicts truss architectures for specified stiffness responses.

Main Results:

• The framework successfully identifies topologically distinct truss lattices matching target anisotropic stiffness.
• It demonstrates the ability to predict architectures for previously unseen stiffness responses.
• Applications include patient-specific bone implants with clinically relevant stiffness.

Conclusions:

• The proposed deep-learning framework effectively addresses the inverse design challenge for architected materials.
• This approach allows for the creation of custom truss lattice structures with tailored stiffness.
• The technology has potential applications in lightweight structures, biomimetic implants, and spatially graded materials.