Abstract: To evaluate the safety of highway bridges due to a random traffic flow, a large number of vehicle-bridge coupling simulations are demanded. However, the efficiency of current related simulation methods is generally limited, which causes the necessary analysis for the safety evaluation is time-consuming and cannot provide an assessment result in time. Therefore, based on the traditional analysis theory for vehicle-bridge coupling vibrations, this paper proposed an approach accelerated by Bi-directional Long Short-Term Memory (BiLSTM) algorithm. This method uses the initial period of bridge response to directly map out the subsequent bridge response, which avoids parts of the iterations needed in traditional calculation methods, leading to the improvement of the overall calculation efficiency. Then, four common bridges were selected as objects and some improved evaluation indicators such as the weighted mean absolute percentage error (WMAPE) and the weighted coefficient of determination (WR²) were proposed specifically. The accuracy, robustness, and calculation efficiency of the proposed method were analyzed in comparison with the traditional simulation method. The results show that the proposed method has a good robustness which attains the high accuracy for predicting the bending moment, the shearing force and, the deflection of bridges under different random traffic densities or under different pavement roughness conditions. Compared with the traditional method, this method can improve the calculation efficiency by 37.98% on average as its WMAPE less than 3.2%, peak absolute error (PAE) less than 2.9%, and WR² larger than 0.98. This indicates that this method could greatly improve the calculation efficiency of the stochastic traffic-bridge coupling vibration with high accuracy and has the potential for the rapid analysis and evaluation of bridge structures.