Leveraging large language models as synthetic experts: A methodology and validation framework for pavement performance assessment

Despite recent advances in automated sensing and analytics, human experts remain indispensable for effective pavement management. However, traditional expert-driven evaluations are frequently constrained by high operational costs, logistical barriers, and inter-rater variability. Alternatively, Large Language Models (LLMs) have demonstrated utility in simulating human responses in social science research. However, their capacity to replicate specialized reasoning in technical engineering domains remains insufficiently characterized. This study investigates the deployment of LLMs as synthetic domain experts to generate input data for pavement performance modeling. A structured questionnaire was developed to quantify the multi-dimensional impacts of surface and subsurface asphalt pavement defects across five performance criteria: safety, user costs, rideability, deterioration rate, and environmental impact. To simulate a diverse professional cohort, 360 synthetic expert profiles were constructed by systematically varying geographic location, organizational affiliation, and area of expertise. Three LLM architectures, Gemini 2.5 Pro, Flash, and Lite, were prompted using these profiles to generate a total of 1,080 survey responses. A multi-dimensional validation framework evaluated response quality, benchmarking synthetic outputs against responses from 12 domain experts and 11 engineering students. The result shows that all models achieved 100% completeness and high internal consistency (Cronbach's Alpha >0.85), with Pro and Flash demonstrating higher geographic awareness and persona-influenced patterns. However, multiple limitations persist The models systematically conflated normative recommendations with established practices and, in some instances, exhibited latent biases that were resistant to persona-based prompting. These findings establish that LLMs can serve as a scalable supplementary tool for structured data generation. However, deploying LLMs as autonomous agents in infrastructure decision-making warrants further investigation.