From AI to Application—Generative AI for Rapid and Efficient Mechanical Design of Composites

The distribution of material phases is crucial to determine the composite’s mechanical properties. Studying the complete structure-mechanics correlation of meticulously ordered material distributions is feasible within a finite number of cases. However, this relationship becomes challenging to discern for complex irregular distributions, hindering the design of material structures that meet specific mechanical requirements. Generative artificial intelligence (AI) is shown to be a useful tool to automatically learn from the existing information and generate new information based on their connections, but its usage for quantitative mechanical research is less understood.<br/><br/>In this work, we aim to combine the cutting edge generative artificial intelligence (GenAI), specifically Stable Diffusion, with Molecular Dynamics simulations and insightful mechanical analysis to design the material distribution within a composite material for optimal mechanical functions. We develop a fine-tuned SD model that generates the matrix-reinforcement material distributions within a rectangular composite sample and provides its corresponding stress fields in uniaxial deformation with accuracy. We use mechanical analysis to extract the composite mechanical properties from the material distribution and stress fields and use variational auto-encoder to reveal the latent space of the mechanical functions for the composite design, enabling its function-based optimization and design. Our findings demonstrate that GenAI can effectively learn critical features from a relatively small training dataset and, by exploring the design space, can accurately and extensively generate composite material distributions along with their corresponding stress fields under load. We also emphasize that this technique is efficient in generating extensive composite designs with valuable mechanical information that determines the stiffness, toughness, and robustness of the material using a single model, a process that would typically require multiple experimental or simulation tests.<br/>We extend this framework by enabling the understanding of the natural language descriptions of the sample geometry, boundary conditions, and loading conditions and validate the prediction of the optimal material distribution with practical experiments using a multi-material 3D printer<br/><br/>Our research framework will enable the efficient design of complex composites with natural language description instead of complex numerical modeling, data-hungry learning, and sophisticated optimization. It will significantly reduce the modeling effort, and the predicted outcome can be directly applied to composite synthesis for validation or application to broad engineering fields that heavily depend on composite materials.