Reservoir model history matching is critical for understanding subsurface uncertainties in rock properties. However, traditional history-matching methods often require numerous forward model evaluations and are sensitive to the initial guess of uncertain model parameters, making the process computationally intensive and potentially unstable. To tackle these issues, we resort to deep learning (DL) technologies for their universal approximation capability in both forward and inverse modeling based on automatic differentiation. In this study, we develop a deep neural network–based history-matching (DNN-HM) workflow as a deterministic approach to enhance the accuracy and efficiency of history matching. The workflow couples two specialized networks: a DL-based forward surrogate model  for fast prediction of multiphase flow and an inference network  for history matching based on prior knowledge and the pretrained ⁠. We assess the performance of the DNN-HM workflow on 2D and 3D two-phase waterflooding problems in heterogeneous reservoirs. After training,  accurately predicts well grid pressures  and saturation ⁠. Starting from a homogeneous prior,  successfully infers a heterogeneous permeability field with low relative error and enables accurate forecasting of production rates (⁠⁠, ⁠), well bottomhole pressures ⁠, and saturation plume propagation ⁠. Sensitivity analysis shows that using longer observational periods improves history-matching accuracy, and the DNN-HM workflow demonstrates strong robustness to observational data noise. Compared to traditional gradient-based methods, DNN-HM achieves higher efficiency, offers transfer learning capabilities, and improves permeability estimation accuracy. Finally, the workflow is extended to 3D cases, demonstrating its scalability and applicability to realistic reservoir scenarios.