The fiber waviness is inevitable in non-crimp fabric (NCF) reinforced composites. It is very challenging to accurately and efficiently predict the material behavior with fiber waviness. This work presents a machine learning approach to the prediction of material behavior of NCF composites under a compressive load. The out-of-plane fiber orientations are first extracted from micrographs of NCF laminates. A digital twinning process is followed to create finite element (FE) models with elementwise fiber orientations. Based on the FE models, a physics-based damage model is employed to generate high-fidelity simulation datasets, capturing the kink-band due to the fiber waviness. With the simulation datasets, convolutional neural network (CNN) models are developed to take the images of the fiber orientations and predict the corresponding stiffness, strength, and stress-strain curves of the NCF composites. The results show that the CNN models can capture spatial information of the fiber orientation and efficiently predict the corresponding material behavior with a high accuracy. In addition, the correlations of the fiber orientations and the final material behaviors are investigated based on the developed CNN models.