Kalman filter-based methods have had remarkable success in the oil industry in recent years, especially to solve several real-life history matching problems. However, as the formulation of these methods is based on the assumptions of gaussianity and linearity, their performance is severely degraded when a priori geology is described in terms of complex distributions (e.g., facies models). The current trend in solutions for the history matching problem is to take into account more realistic reservoir models, with complex geology. Thus the geological facies modeling plays an important role in the characterization of reservoirs as a way of reproducing important patterns of heterogeneity and to facilitate the modeling of the reservoir rocks petrophysical properties. This thesis introduces a new methodology to perform the history matching of complex geological models. This methodology consists of the integration of Kalman filter-based methods, particularly the method known in the literature as Ensemble Smoother with Multiple Data Assimilation (ES-MDA), with a parameterization of the geological facies through techniques based on deep learning in autoencoder type architectures. An autoencoder always consists of two parts, the encoder (recognition model) and the decoder (generator model). The procedure begins with the training of a set of facies realizations via deep generative models, through which the main characteristics of geological facies images are identified, allowing for the creation of new realizations with the same characteristics of the training base, with a low dimention parametrization of the facies models at the output of the encoder. This parameterization is regularized at the encoder to provide Gaussian distribution models in the output, which is then used to update the models according to the observed data of the reservoir through the ES-MDA method. In the end, the updated models are reconstructed through deep learning (decoder), with the objective of obtaining final models that present characteristics similar to those of the training base. The results, in three case studies with 2 and 3 facies, show that the parameterization of facies models based on deep learning can reconstruct facies models with an error lower than 0.3 percent. The proposed methodology generates final geological models that preserve the a priori geological description of the reservoir (facies with curvilinear channels), besides being consistent with the adjustment of the observed data of the reservoir.