Generative models struggle with kirigami metamaterials.

Area of Science:

• Materials Science
• Machine Learning
• Computational Design

Background:

• Generative machine learning models excel at designing metamaterials with target properties.
• Metamaterials' behavior depends on their internal structure, making design complex.
• Kirigami metamaterials introduce design restrictions due to interdependent cuts.

Purpose of the Study:

• To investigate potential survivorship bias in generative models for metamaterials.
• To evaluate the performance of leading generative models on kirigami structures.
• To identify limitations of current generative models in creating diverse kirigami metamaterials.

Main Methods:

• Assessed four popular generative models: Variational Autoencoder (VAE), Generative Adversarial Network (GAN), Wasserstein GAN (WGAN), and Denoising Diffusion Probabilistic Model (DDPM).
• Focused on generating kirigami structures with constraints on cut intersections.
• Analyzed the suitability of similarity measures, specifically Euclidean distance, for kirigami geometries.

Main Results:

• Perceived success of generative models in metamaterial design may be influenced by survivorship bias.
• Prohibiting cut intersections in kirigami complicates similarity measure identification.
• VAE and WGAN models are significantly impacted due to their reliance on Euclidean distance, which is unsuitable for kirigami geometries.

Conclusions:

• Current generative models face significant limitations when applied to kirigami metamaterials.
• The effectiveness of models like VAE and WGAN is compromised by unsuitable similarity metrics for complex geometries.
• Further research is needed to develop robust generative approaches for diverse kirigami metamaterial design.