Abstract: Supervised deep learning inversion relies on labels of subsurface media for training. However, the measured data from electrical resistivity tomography usually lack such labels. Supervised learning based on significant synthetic data generated through forward modeling fails to obtain reliable inversion results for complex data outside the training set. This study proposed a semi-supervised learning inversion method based on forward modeling constraints, combining with labeled and unlabeled data for training, to enhance the inversion accuracy and the capability to characterize complex geological structures. The inversion results of synthetic data demonstrate the strong generalization capability of the proposed method. The inversion results of measured data indicate that compared to conventional least-squares inversion, semi-supervised inversion provides higher resolution and particularly more accurate characterization of vertical lithological boundaries. Overall, the proposed method offers a novel approach for the inversion of the data derived from high-resolution electrical resistivity tomography.