In this paper, we propose a multiscale data-driven framework for Fiber Reinforced Polymer (FRP) composites. At the mesoscopic scale, the 3D stress–strain material database is collected by the multilevel computational homogenization (FE²), in which the Representative Volume Elements (RVEs) are generated through the X-ray microtomography (Micro-CT) aided technique. Such so-called “material twin” technique is able to reproduce the high quality yet operational geometric mesoscopic details of FRP composites. At the macroscopic scale, the distance-minimizing data-driven approach is adopted to simulate the structural behavior by directly searching the material database without employing the constitutive model. This data-driven approach can significantly save the computational cost because the needed mesoscopic behaviors are previously prepared by the offline mode. The numerical results demonstrated that the proposed data-driven framework is a promising scheme. From material database collection to data-driven analysis, this framework opens a new channel for FRP simulation in the data science paradigm.