Benchmarking Large Language Models for Embodied Carbon Calculations of Structural Systems

The increasing accessibility of open Artificial Intelligence (AI) platforms and conversational agents has encouraged their adoption across engineering disciplines, including sustainability assessment. This study investigates the capability of Large Language Models (LLMs) to estimate the embodied carbon of structural systems and provides a comparative benchmark of their performance. Five leading AI models—ChatGPT, Gemini, Claude, DeepSeek, and Grok—were evaluated using three structural case studies. Each model was first tested without task-specific training and subsequently reassessed after being provided with a reference example, enabling evaluation of both baseline accuracy and learning potential. Results show that all models were generally effective in determining material quantities; however, notable variations were observed in the selection and application of embodied carbon factors (ECFs). Among the untrained models, ChatGPT 5 achieved the closest agreement with benchmark values, exhibiting a deviation of only +4.3%, whereas Claude Sonnet 4.5 produced the largest discrepancy at +93.4%. Following reference-based training, substantial improvements were observed across all models, with embodied carbon estimates for simple structural systems converging to within ±2% of benchmark results. The findings demonstrate the potential of LLMs to support rapid and efficient sustainability assessments in structural engineering while highlighting the importance of appropriate guidance and validation to ensure reliable outcomes. This work provides insights into the strengths, limitations, and future prospects of AI-assisted embodied carbon assessment in sustainable design practice.