Abstract: The classical stepwise integration method is usually used to solve the dynamic response of structures, but the accuracy and stability will be reduced if the time step is too large. If the time step is too small, the calculation efficiency will be particularly low. Machine learning can significantly improve learning efficiency, but traditional machine learning methods require a large amount of training data to perform structural dynamic response calculations. In order to improve the interpretability and generalization ability of data-driven machine learning methods, this paper adopts a physics-informed neural network (PINN) approach to solve structural dynamic response analysis, and introduces the structural dynamic response equation into the loss function as a physical constraint to endow the neural network with physical meaning. In essence, the method converts the problem of directly solving the dynamic differential equation into the optimization problem of the loss function to obtain the solution of the ordinary differential equation. Subsequently, the feasibility and effectiveness of the method proposed are verified by a numerical analysis of the dynamic response of the structure under different excitations, and the influence of different neural network parameters on the convergence of the PINN model is discussed. Numerical results show that compared with the data-driven machine learning method, the PINN method can effectively analyze the structural dynamic response without any system response data. Compared with the classical stepwise integration method, the method proposed belongs to the spatial domain optimization process, which is independent of the time domain, thusly it is not affected by the time step.