
This study proposes an AI-assisted wave-based inverse reconstruction framework for characterizing spatially varying fiber orientation fields in thin-walled short-fiber-reinforced composite plates. The problem is formulated as an inverse mapping from time-domain signals collected by multiple sensors to a discretized fiber orientation field, enabling direct prediction without iterative inversion. To generate training data efficiently, several homogenized finite element models are developed and compared with a detailed finite element model that explicitly resolves the microstructure. The results show that all homogenized models provide substantial reductions in computational cost, while their accuracy differs. Among them, the unit-cell-based reinforcement-part averaging model achieves the most favorable balance between accuracy, efficiency, and robustness, and is therefore selected for large-scale forward simulations. Based on the generated dataset, a multilayer perceptron is trained to predict the fiber orientation field from boundary wave signals. A double-angle representation is adopted to account for the periodic nature of orientation. The results demonstrate that the model can accurately reconstruct both random and structured orientation distributions, with stable training behavior and good generalization performance.