Machine learning-assisted multiscale optimization for continuous fiber reinforced composites

Continuous fiber reinforced composites (CFRCs) are key material systems in fields such as automotive and aerospace. Recently, additive manufacturing technology has provided a new methods for the controlled preparation of CFRCs. However, the material anisotropy and nonlinear properties caused by the non-uniform spatial distribution of fiber orientation and microstructural features pose significant challenges in multiscale modeling and concurrent optimization. In this paper, a neural network-assisted multiscale concurrent optimization (NNMCO) algorithm for the continuous fiber orientation and macrostructure topology of anisotropic composites is proposed. In order to do this, firstly, a mapping relationship between the micro fiber orientation and effective material properties of representative volume element (RVE) is constructed using fully-connected neural network (FCNN). Then, at the macroscale, the density-based Solid Isotropic Material with Penalization (SIMP) method is used to optimize the macrostructure topology by penalizing the stiffness of intermediate-density elements. Meanwhile, at the microscale, the fiber orientation is optimized during the iteration process according to the principal strain alignment (PSA) method to maximize local stiffness. The Gaussian filtering smoothing technique was used to smooth the local fiber distribution and avoid getting trapped in local optima, and the streamline algorithms were employed to generate smooth, continuous fiber paths. Finally, the efficiency and applicability of the developed method are further confirmed via 2D/3D numerical examples, 3D printing preparation, load-displacement experiment, and digital image correlation (DIC) testing. A concurrent multiscale optimization strategy is introduced for CFRCs fabricated via 3D printing.