Opinion dynamics models are widely used to understand how opinions spread, evolve, and converge within social systems. However, applying these models to real-world data presents a significant challenge: the calibration of empirical observations to fit the underlying theoretical framework. In this work, we introduce a graph neural network-based approach for estimating the parameters in bounded confidence-type opinion dynamics models. In this simulation-based inference framework, graph neural networks are used to encode and summarize the structural and opinion-related features of the data, and a coupling flow network is trained to infer the parameters governing the opinion dynamics model. To enhance the efficiency and scalability of our method, we employ graph sampling techniques that simplify the network structure before passing it to the neural network. We evaluate our approach on synthetic networks and further validate its performance using empirical datasets, demonstrating its effectiveness in recovering meaningful model parameters. We have done comparisons against likelihood methods, and the results indicate the efficacy and efficiency of the proposed model.