Creating Synthetic Road Networks Using Large Language Models: An Iterative Calibration Framework

Source abstract

Digital twins are used to represent and analyse urban environments, providing a basis for simulation, scenario testing, and communication among stakeholders. While effective, such models are typically developed as unique, data‐intensive representations of specific cities. Synthetic cities extend this paradigm by enabling the systematic generation of entire families of digital twins. They provide controlled, repeatable, and easily calibratable environments, making it possible to conduct analyses that are more general, reproducible, and independent of costly or sensitive field data. Generating synthetic cities addresses several key needs in urban research. It supports the design and evaluation of regulatory scenarios, facilitates climate adaptation analyses, and enriches participatory design by providing semantically meaningful models for discussion. By decoupling experimentation from real‐world data constraints, synthetic cities provide a robust framework for testing alternative urban forms and policies while preserving privacy and reducing costs. This article introduces a novel method for generating synthetic cities using large language models (LLMs). Specifically, we propose an iterative LLM‐guided framework in which the model directly generates high‐level synthetic road structures and calibrates them through tool‐based structural feedback. Unlike procedural generators and neural generative models, which usually require manual parameter tuning and offer limited ways to enforce high‐level semantic constraints, our LLM‐based framework translates natural‐language requirements into quantitative specifications and validates generated networks against target metrics. This supports the creation of adaptable, semantically meaningful synthetic cities suitable for simulation and scenario analysis.