Wang, C.G., Zhou, Y.J., He, X.J., Huang, X.Y. and Lu, F., 2022. Full-waveform inversion based on deep learning and the temporal modified and spatial optimized symplectic partitioned Runge-Kutta method. Journal of Seismic Exploration, 31: 501-521. This study uses deep learning techniques to propose a full-waveform inversion (FWI) method. This method uses the Temporal Modified and Spatial Optimized Symplectic Partitioned Runge-Kutta (TMSOS) method, for the forward modeling in the FWI process. Additionally, optimizer, loss function in deep learning are utilized to perform FWI. We used recurrent neural networks in deep learning to implement the forwarding modeling method-TMSOS. This method uses the second-order MSPRK scheme and an eighth-order optimized finite difference scheme for temporal and spatial discretization, respectively, to obtain high precision with lesser computational effort. The Nadam optimizer, packaged in the Tensorflow software, is used to optimize the model parameters in this study. Additionally, Huber loss is used as the objective function of FWI. Several numerical simulations were done to verify the effectiveness of the proposed method. First, the effectiveness of the TMSOS forward modeling method was compared with the finite difference (FD) method. Then FWI was performed using the TMSOS forward modeling method for the velocity anomaly model, the third-order model, and the complex Sigsbee velocity model. Finally, the inversion results under the three loss functions were compared. The numerical results suggest that the TMSOS forward modeling method has better computational efficiency and smaller resultant numerical dispersion than the traditional FD method. The inversion results of the TMSOS method are closer to those of the actual model, with a better inversion effect. Additionally, the Huber loss function has better stability in selecting the learning rate.