Based on the combination ofdata- and model-driven approaches, this study expanded the labels of the training set through model inversion results, and added the model inversion objective function to the deep learning algorithm. By constructing a new loss function, this study proposed a seismic impedance optimization inversion method combining model inversion with deep learning inversion. The semi-supervised deep learning network inversion under a pseudo-label was achieved using the RNN network structure. The network inversion results were used as the initial model to participate in the model inversion. The final optimization inversion was completed by continuous iterative optimization of both network and model inversion. The method proposed in this study proves to possess high inversion accuracy and practicability, as demonstrated by the synthesis of the Marmousi model and the actual data.