Abstract: In this paper, a dynamic adaptive model optimization method based on GM (1, 1) model is proposed by combining the grey system theory with gradient descent algorithm, which is used for bridge seismic damage prediction. Combined with the characteristics of strong randomness of ground motion and bounded structural damage in the study of bridge seismic damage, the GM (1, 1) model is optimized, and the gradient descent algorithm is introduced to realize the dynamic optimization of parameters. The feasibility and accuracy of the model optimization method are verified by establishing the finite element model of a four span pre-stressed concrete composite box girder bridge and using the measured ground motion records for nonlinear dynamic time history analysis. The results show that the optimized model only needs 5~6 initial data to effectively predict the seismic vulnerability of bridges, the maximum error is controlled within 7%, and the prediction accuracy of minor damage is the highest. This study provides a lightweight prediction scheme for earthquake prone areas, and multi intensity vulnerability assessment can be carried out based on a small amount of data.