Deciphering AI behaviors in microstructural material modeling of composites using diffusion models

Micro-structural modeling of representative volume elements (RVEs) of composite materials with subsequent computational homogenization is key for developing and establishing accurate and reliable material models on the macroscopic and the structural levels. Such methods require RVEs that represent the real microstructure of the material. Recent advancements in generative AI models have demonstrated their potential in high resolution image synthesis for micro-heterogeneous materials such as fiber-reinforced used in wind turbine blades. This research explores the application of diffusion models to generate statistically representative microstructures of the aforementioned material. Hallucinations in generative models are a unique challenge. In associative networks, a problem of exact retrieval of specific data points transforms into a desired solution in generative modeling where merging individual data points into coherent, larger structures is preferred. This inversion of priorities underscores the complex interplay between memorization and generalization, which this study seeks to explore in the context of materials science. Collapse, where a model’s output variety diminishes as it shifts from generalization to memorization, is also investigated. Moreover, the study explores if models exhibit a gradual reduction in output diversity, aligning with the theoretical concept of dimensional collapse observed in statistical physics, which can inadvertently lead to bias reinforcement. This particularly holds in underrepresented data classes. This study investigates these behaviors and contextualizes them within the application of AI in materials science.