A novel PCE-NN-Transfer-Learning framework was developed to infer physical behavior in the presence of multiscale interactions. We specifically address inferences relevant to delamination and associated constitutive laws and judiciously introduce cross-scale information for a carbon-reinforced composite material. Representative volume methods (RVE) with periodic boundary conditions and random material and geometric properties augment the microscale system. Specifically, Beta probability distributions are attached to Young’s modulus, Poisson’s ratio, and maximum strain and stress of tows and resins, cohesive strength of the interface, the shape and volume fraction of the ellipse tows. The microscale strain-material-properties-stress dataset is obtained with quadrature failure observed and analyzed. Polynomial chaos expansions (PCE) with principal components analysis for smoothing and dimension reduction are used to obtain the stiffness matrix, maximum strain, and stress of the whole RVE trained on this dataset, providing a transition between microscale and mesoscale descriptions. Each mesoscale stiffness and maximum strain and stress realization is then provided as input to a user-defined material subroutine in Ls-Dyna to perform bending simulations, thus yielding the macroscale bending force-displacement dataset used for training and validation. First, a neural network is trained on the micro-to-macro scale dataset: micro-input layer-wise material properties and macro-output stiffness of the bending force-displacement; it is then retrained on its unfrozen layers on the meso-input, PCE-predicted stiffness matrix, maximum strain and stress, to the same macro-output. Finally, a PCE-transfer-learning cross-scale framework for multilayer-stacked carbon-fiber-reinforced systems is constructed, from micro (RVE) scale to meso (stiffness, maximum strain, and stress) to macro (bending stiffness) to analyze complex composites with high accuracy and low computational efforts.