Abstract Manager

In the domain of civil engineering, optimizing bridge configurations for low-volume road bridges (LVRBs)—structures typically serving fewer than 400 vehicles per day—presents unique challenges due to their large number, distinct risk profiles, and sometimes remote locations. This study explores AI-driven tools for structural system optimization, adapting rule-based expert systems to develop user-friendly prompts for large language models (LLMs), thereby enhancing accessibility and efficiency in design processes.

The investigation begins with fundamental LVRB parameters, such as hydraulic openings, riverbed and approach elevations, and geotechnical conditions to determine optimal substructure and span combinations. In parallel, risk assessment tables are integrated as augmented datasets to evaluate the impacts of various design exceptions, enabling more cost-effective configurations.

This work advocates hybrid methodologies for rapid prototyping that combine expert systems for rigorous inference with LLM prompts. Recognizing LLMs' susceptibility to hallucinations, expert systems serve as essential verifiers to validate prompt-generated outputs, ensuring accuracy and code compliance. Furthermore, leveraging expert systems to craft prompts yields well-structured formats, such as Chain-of-Thought (CoT) or Tree-of-Thoughts (ToT), which maximize LLM reasoning capabilities and mitigate hallucinations through guided, step-by-step logic.

Preliminary findings demonstrate that LLMs can reason through project constraint logic and can produce reliable preliminary designs at the element level when well structured detailed prompts are provided.