As e-commerce growth drives customer expectations for faster and more flexible delivery, companies increasingly rely on distributed fulfillment networks to shorten the time between order placement and delivery. Evaluating the cost and flow implications of these networks typically requires high-fidelity simulation models that solve individual order-level fulfillment decisions across all feasible routes. The process is computationally intensive and limits rapid scenario-based analyses for strategic and tactical planning. This project develops a graph neural network (GNN) surrogate model that approximates the aggregate outputs of such a simulator, specifically total episode shipping cost and route-level fulfillment flows, without explicitly solving the optimization problem for every order. The final model achieves a testset nonzero flow F1 of 0.860, and a weighted MAPE of 1.39% on episode total cost across 150 held-out episodes. The trained model produces a per-episode runtime that is approximately 40 times faster than the high-fidelity simulator. These results demonstrate that a simplified, machine learning-driven network flow model can serve as a practical tool for rapid what-if analysis across diverse demand and inventory scenarios. It can enable supply chain planners to evaluate network design alternatives at a fraction of the computational cost of traditional simulation approaches.
