Scalable data-driven micromechanics model trained with pairwise fiber data for composite materials with randomly distributed fibers

A machine learning (ML) model can provide a precise prediction very quickly, if it is well trained with a massive amount of reliable training data. A finite element method (FEM) is often employed to generate substantial training data. However, such a training process can be computationally burdensome especially for a geometrically complex structure. More critically, a specific size and/or configuration of a training model may confine the applicability of the trained model to the same kind only. In this study, we present a scalable ML approach with an efficient training strategy for micromechanical analysis of fiber-reinforced composite materials. Here, a scalable data-driven micromechanics model (SDMM) is proposed for predicting stresses in unidirectional composites with random fiber arrays. The training data for SDMM is defined in the unit of a fiber pair. A single dataset is composed of a stress value between a fiber pair and an image highlighting the pair with nearby fibers affecting the stress. Therefore, the training microstructures can be considerably small, but the pairwise ML model can be applied to every pair of adjacent two fibers inside a much larger microstructure. The scalability of SDMM is demonstrated by predicting the maximum principal stress values acting between every fiber pair in a super-sized representative volume element. The accuracy of the prediction results is evaluated by finite element analysis results. It is shown that a certain number of nearby fibers is required in the training datasets for accurate prediction.