Abstract Hydraulic fracturing is the key enabling technology for the economic development of unconventional reservoirs, where well productivity depends on the creation of an extensive, conductive fracture network. Accurate characterization of effective fracture geometry and reservoir properties is essential for predicting well performance and optimizing stimulation design. However, conventional historymatching approaches in these complex systems are hindered by high-dimensional uncertainty, strong nonlinearity, and the high computational cost of large-scale simulations. This study presents an integrated workflow that couples a GPU-accelerated Embedded Discrete Fracture Model (EDFM) with the Ensemble Smoother with Multiple Data Assimilation (ESMDA) and AI-based surrogate modeling to perform automatic, uncertainty-aware history matching in unconventional reservoirs. The EDFM provides high-fidelity representation of complex hydraulic and natural fractures, GPU acceleration enables rapid evaluation of large ensembles, and AI surrogates reduce the number of full-physics simulations required per assimilation cycle. The workflow systematically assimilates production and pressure data to estimate both static and time-dependent parameters, while quantifying uncertainty through posterior parameter distributions. The approach is validated using two benchmark cases: (1) a single well with a single hydraulic fracture and time-dependent fracture permeability decline, and (2) a multi-layer, three-well configuration with multiple uncertain parameters, including fracture and matrix properties. In both cases, the workflow achieved accurate history matches and recovered the true parameter values with relative errors typically below 15%, while requiring only seconds of wall-clock time per simulation on a GPU. Results demonstrate that the proposed EDFM-ESMDA-AI framework enables scalable, automated, and physically consistent dynamic reservoir characterization, offering a practical solution for field-scale unconventional reservoir modeling.