Abstract: Vertical heterogeneity greatly influences fluid flow and oil recovery. Traditional methods, such as 4-D seismic testing and vertical interference tests, are costly and time-intensive, while well logging provides limited spatial coverage. This research explores deep learning as a data-driven alternative, using well history data and production rates to estimate stylolite zone characteristics and permeability distributions. We implemented neural network architectures, including CNNLSTM, U-Net, and transformers encoder, to predict the size and location of fluid pathways within stylolite zones. To address real-world complexities, we moved forward to the permeability characterization of the reservoir. We developed deep learning architectures, including transformer encoder, and text-to-image diffusion model, to conduct detailed reservoir characterization with fewer and simpler inputs.