Image-driven machine learning constitutive mapping the confluence of material and micro-topological attributes in short fiber polymer composites

Short Fiber Reinforced Polymer Composites have attracted significant attention due to their customizable mechanical behavior. However, the inherent complexity and heterogeneity of these composites pose critical challenges in accurately quantifying their constitutive properties and optimizing micro-structural designs through experimental and computational methods. To circumvent the intricacies and labour-intensive nature of traditional experimental and numerical modelling approaches, we propose a machine learning (ML) model that takes the microstructural topology of fiber distribution as input in the direct image form, along with material properties of fibers and matrix, and generates insightful visualization of the stress and strain fields as output. Such direct image-based mapping of the topological architectures significantly alleviates the hurdle of numerical representation of randomly oriented and randomly located fibers along with intrinsic material properties of the constituting elements, while making the ML model capable of handling high-dimensional datasets that would be impossible to deal with based on traditional data-driven approaches. We have employed finite element simulations based on Representative Area Element to train a paired image-to-image (pix2pix) model exploiting conditional Generative Adversarial Networks, considering three scenarios of constitutive mapping in the order of increasing complexity, (1) only varying material properties, (2) only varying microstructural geometry, and (3) simultaneously varying material properties and microstructural geometry. Developing an efficient, strictly image-based ML algorithm for constitutive mapping of composites can reduce analysis time and cost, while enabling qualitative and quantitative insights for microstructure development, multiscale design, and multifunctional optimization. Further, it establishes the foundation for image-to-video learning of evolving heterogeneous composite mechanics.