The microstructural heterogeneity affects the macroscale behavior of materials. Furthermore, the loading at the macroscale changes the microstructural morphology. These structure-property relations are often modeled using multiscale approaches such as the finite element method (FEM). However, multiscale FEM requires many calculations at the local scale which are often infeasible to tract. Here, we report a significantly faster data-driven machine learning based approach in multiscale materials modeling. In this work, we first develop a deep learning model to predict stress tensor field at the local level in a fiber-reinforced composite material. A mapping between the spatial arrangement of fibers and the corresponding 2D stress tensor field is achieved by using a convolutional neural network (CNN), specifically a U-Net architecture. Then we use this model in a multiscale simulation framework with excellent accuracy in homogenization of elastic material properties and localization of stress tensor field. Our approach shows tremendous potential in the multiscale analysis of materials.