Accelerated multiscale mechanics modeling in a deep learning framework

Abstract

Microstructural heterogeneity affects the macro-scale behavior of materials. Conversely, load distribution at the macro-scale changes the microstructural response. These up-scaling and down-scaling relations are often modeled using multiscale finite element (FE) approaches such as FE-squared ($FE^2$). However, $FE^2$ requires numerous calculations at the micro-scale, which often renders this approach intractable. This paper reports an enormously faster machine learning (ML) based approach for multiscale mechanics modeling. The proposed ML-driven multiscale analysis approach uses an ML-model that predicts the local stress tensor fields in a linear elastic fiber-reinforced composite microstructure. This ML-model, specifically a U-Net deep convolutional neural network (CNN), is trained separately to perform the mapping between the spatial arrangement of fibers and the corresponding 2D stress tensor fields. This ML-model provides effective elastic material properties for up-scaling and local stress tensor fields for subsequent down-scaling in a multiscale analysis framework. Several numerical examples demonstrate a substantial reduction in computational cost using the proposed ML-driven approach when compared with the traditional multiscale modeling approaches such as full-scale FE analysis, and homogenization based $FE^2$ analysis. This approach has tremendous potential in efficient multiscale analysis of complex heterogeneous materials, with applications in uncertainty quantification, design, and optimization.