Abstract: In this study, a methodology is proposed for the lightweight characterization and efficient prediction of the nonlinear dynamic responses in jointed structures. Proper orthogonal decomposition (POD) is employed to reduce the dimensionality of nonlinear vibration responses, whereas an adaptive neural network (ANN) surrogate model is utilized to predict the projection coefficients associated with the POD orthogonal basis. A nonlinear dynamic system with a bolted joint is considered as the research object. Latin hypercube sampling is employed to generate a sample set of design parameters. The alternating frequency-time harmonic balance method (AFT-HBM) is used to generate a snapshot dataset of nonlinear frequency responses. The mapping relationship between the characteristic parameters of the joint, the excitation amplitude, and the projection coefficients is established using the POD-ANN approach. The predictive performance of the model is quantitatively evaluated using the improved Akaike’s information criterion (AIC) and compared with three methods reported in the literature. The results demonstrate that the proposed method achieves efficient prediction of nonlinear vibration responses and is computationally more efficient than existing methods for large-scale sampling problems. In a sampling task with a scale of 10⁴, the computational time is reduced to one-quarter of that required for a single AFT-HBM calculation, and the dimensionality reduction ratio reaches 50, with the error remaining below 1.5%.