A Tool-Augmented LLM Framework for Life-Cycle Carbon Assessment of Bridges

To support carbon neutrality in civil infrastructure, bridge carbon assessment should be integrated into the early design stage, when key design options remain adjustable. However, this stage usually lacks detailed bills of quantities, material sources, transport distances, construction methods, and maintenance schedules, limiting the direct application of conventional life-cycle assessment (LCA). Published bridge carbon cases can provide reference data for early-stage assessment, but differences in system boundaries, functional units, and reporting formats hinder their direct comparison and reuse. To address this issue, this study first used a cable-stayed bridge in Ningbo, China, as a traceable LCA calculation case to establish the accounting procedure, system boundary, and reporting fields. It then combined this case with published bridge carbon cases to develop a standardized bridge carbon reporting database. The resulting database contains 39 cases across six harmonized bridge-type categories and supports similar-case retrieval and early-stage prediction modeling. Linear, polynomial, and random forest regression models were developed using the most consistently available early-stage parameters to provide initial emission estimates under limited design information. Under the fixed 70:30 data split, the random forest model achieved the best internal validation performance, with a test R² of 0.97 and a mean absolute percentage error (MAPE) of 29%. Static carbon-flow and dynamic time-weighted carbon-flow indicators are further introduced as normalized comparative indicators for interpreting stage-level contributions and temporal emission profiles. The results show that temporal weighting can alter the relative ranking and interpretation of bridge-type carbon performance. As applying these analytical modules still requires substantial expert involvement, a tool-augmented large language model (LLM) assistant was further implemented as the framework’s interaction layer to reduce this operational burden. It structures natural-language inputs, invokes predefined database and analytical tools, and explains their outputs, while engineers remain responsible for verifying the results. The proposed framework provides a traceable, human-in-the-loop approach to supporting early-stage bridge carbon assessment under incomplete design information.