The conventional geo-electromagnetic data inversions are mostly based on gradient optimization methods. However, this type of method can only provide a single “optimal” inverse model under specific prior conditions, which cannot effectively evaluate the reliability and uncertainty of the inversion results. The widely used uncertainty quantification (UQ) methods are based on the theory of Bayesian inference. Although they have achieved success in many applications, they suffer from the curse of dimensionality and low efficiency. To overcome these problems, we propose a novel UQ strategy for geo-electromagnetic inversions based on Bayesian processes and surrogate modeling with adaptive deep neural network (DNN). In this method, an embedded DNN is used for forward modeling in the Bayesian inference to improve computational efficiency. The training of the DNN is divided into two stages. First, a predesigned small training set is used and the resulting DNN only gives a low-accuracy result. Second, this DNN is fine-tuned dynamically during the Metropolis-Hastings (M-H) sampling process, in which the training set is adaptively supplemented according to the modeling errors. Compared to the conventional data-driven approach, this dynamically adaptive constructing method of the training set can greatly reduce the training set and constantly maintain high accuracy in forward modeling. We demonstrate the effectiveness and practicality of our surrogate modeling Bayesian and analyze the effects of different sampling numbers, noise levels, prior distributions, and sampling radius. Compared with Occam’s inversion and conventional Bayesian inversions, our method shows good robustness and high accuracy, making it an effective Bayesian inversion technique.