Abstract: Large-scale transport simulation models face the critical challenge of computational complexity. This complexity significantly hampers calibration and model exploration, such as scenario discovery, due to the extensive time required for each simulation run. Traditionally, two approaches have been employed to mitigate this issue: model simplification through spatial and/or temporal aggregation and scope reduction by focusing on specific segments of the transport network or population subsets. Alternatively, analytical or statistical model approximations, known as metamodels, have been utilized. Recently, Machine Learning (ML)-based metamodels have gained popularity for their potential to streamline simulations. However, these models often struggle with out-of-distribution scenarios, failing to accurately represent the original simulation under new conditions, such as policy interventions. This headline presentation will introduce recent advancements in causal metamodeling, an approach that integrates ML-based metamodeling with domain-specific knowledge to produce simulation approximations that not only remain true to the original model but also operate with significantly enhanced speed. Drawing on the recent and ongoing research, I will outline the foundational concepts, share preliminary findings, and explore the challenges and opportunities this research presents. Through causal metamodeling, we aim to significantly improve the scalability and generalization capabilities of simulation models, opening new avenues for comprehensive and efficient scenario analysis and policy advisory.