Traffic signal control is one of the most important challenges for resolving traffic congestion in modern cities. Many previous studies that attempted to solve this problem were based on a simulation optimization approach. However, these studies focused on the optimization procedure without considering two fundamental limitations of simulation: designing a simulation environment and running a simulator during optimization. To overcome these limitations, this study proposes a simulationoptimization framework using a deep-neural-network-based metamodel. The proposed framework can be summarized in five steps: (1) parameterizing real-world features, such as road networks and traffic demands; (2) constructing sample intersections as simulation environments; (3) gathering training data by running a simulator; (4) fitting the metamodel with the data to predict the output of the simulator; and (5) immediately producing the optimized traffic signal via the trained model whenever a new real instance is provided. Based on numerical experiments, we evaluated the performance of the proposed framework in terms of the prediction accuracy of the metamodel and optimized the solution quality. The experiments show that the proposed framework is comparable with the existing simulation optimization framework in terms of solution quality; however, it significantly reduces the computation time. The experiments also verify that its performance is robust for unseen road networks and traffic demands. Moreover, the metamodel can easily accommodate variants of problems with multi-objective functions or additional constraints, such as the minimum duration of pedestrian crosswalks and fixed cycle lengths.