This study focuses on developing novel modeling techniques for fiber-reinforced composites with polymer and ceramic matrix based on Peridynamic approach. To capture the anisotropic material behaviors of composites under quasi-static and dynamic loading conditions, a new peridynamic model for composite laminate and a modified peridynamic approach for non-uniform discretization are proposed in this study. In order to achieve the numerical implementation of the proposed model and approach, a mixed implicit-explicit solver based on GPU parallel computing is developed as well. The new peridynamic model for composite laminates does not have any limitation in fiber orientation, material properties and stacking sequence. It can capture the expected orthotropic material properties and coupling behaviors in laminates with symmetric and asymmetric layups. Unlike the previous models, the new model enables the evaluation of stress and strain fields in each ply of the laminate. Therefore, it permits the use of existing stress- or strain-based failure criteria for damage prediction. The computation of strain energy stored at material points allows the energy-based failure criteria required for delamination propagation and fatigue crack growth. The capability of this approach is verified against benchmark solutions, and validated by comparison with the available experimental results for three laminate layups with an open hole under tension and compression. The modified peridynamic approach for non-uniform discretization enables computational efficiency and removes the effect of geometric truncations in the simulation. This approach is a modification to the original peridynamic theory by splitting the strain energy associated with an interaction between two material points according to the volumetric ratio arising from the presence of non-uniform discretization and variable horizon. It also removes the requirement for correction of peridynamic material parameters due to surface effects. The accuracy of this approach is verified against the benchmark solutions, and demonstrated by considering cracking in nuclear fuel pellet subjected to a thermal load with non-uniform discretizations. Unlike the previous peridynamic simulations which primarily employs explicit algorithm, this study introduces implicit algorithm to achieve peridynamic simulation under quasi-static loading condition. The Preconditioned Conjugate Gradient (PCG) and Generalized Minimal Residual (GMRES) algorithms are implemented with GPU parallel computing technology. Circulant preconditioner provides significant acceleration in the convergence of peridynamic analyses. To predict damage evolution, the simulation is continued with standard explicit algorithms. The validity and performance of this mixed implicit-explicit solver is established and demonstrated with benchmark tests.