Graph Neural Network-based Simulators (GNS) provide accurate and computationally efficient predictions of granular flow dynamics, making them suitable surrogates for inverse design and control problems. However, training these models for new materials remains a significant computational bottleneck. We propose a transfer learning framework to efficiently adapt pre-trained GNS models to untrained materials. Our framework integrates three complementary innovations: (1) a data curation strategy that selects samples with maximum information content by quantifying distributional shifts between source and target domains; (2) a scheduled sampling technique that gradually transitions from teacher-forced to autoregressive prediction during training, significantly enhancing multi-step stability; and (3) a selective fine-tuning approach that updates only material-sensitive neural network layers while preserving generalized physics knowledge. We validate our approach on the challenging task of adapting GNS models across varying internal friction angles in granular flows, using training data generated by the Material Point Method simulating 2D granular column collapse under gravity in rectangular geometry. Fine-tuning the pre-trained model with the proposed framework achieves a tenfold reduction in test MSE, while requiring only 0.25% of the original training cost. Compared to naive fine-tuning, it attains comparable accuracy using only 50% of the target data and 10% of the model parameters. This framework holds promise for applications in geotechnical engineering, landslide prediction, and real-time control of granular flow systems, where rapid adaptation to different material properties is crucial.