LLM-Assisted Screening Method for Large-Scale Transportation Model Calibration

Recently, disaggregated models have surpassed aggregate approaches in reflecting real-world travel behaviors. Among them, Activity-Based Models (ABMs) simulate person-level daily activities across a wide span of the transportation system, including emerging models such as electric vehicles and shared mobility. With thousands of parameters and a lack of closed-form structure, ABMs present significant calibration challenges, including high-dimensional complexity, conflicting objectives, and prohibitive computational costs. While current Bayesian Optimization (BO) methods address these issues through dimension reduction, they often sacrifice accuracy by relying on restrictive sparsity assumptions. Furthermore, existing approaches frequently under-utilize transportation domain knowledge and its modular structures, limiting their robustness and scalability. This study proposes a large language model (LLM) assisted BO framework that leverages parameter roles to prioritize influential variables, thereby reducing computational overhead and enhancing results. The optimization is further refined by an entropy-based acquisition function designed to mitigate output saturation from extreme inputs, effectively narrowing the search space and avoiding local optima. By exploiting the inherent modularity of ABMs, the framework implements a sequential calibration workflow tailored specifically for multi-modal transportation models. Experiments show that the proposed method achieves lower evaluation cost and higher calibration accuracy than state-of-the-art alternatives. The framework is applicable and computationally scalable, extending beyond ABMs to other large-scale problems in transportation and related fields. The ABMs' outputs provide a rich baseline resource for policy makers, operators, and other upstream models.