According to research published on arxiv.org, a multi-agent framework for automated reinforced concrete highway barrier design achieved a 98.3% compliance rate with engineering specifications, compared to just 11.7% for the best-performing standalone large language model.
The study addresses limitations in using LLMs directly for structural engineering design, including “hallucination, numerical reasoning errors, and insufficient integration with physics-based analysis,” according to the paper. The framework implements a “generation-validation-modification” closed-loop system using AutoGen’s multi-agent orchestration capabilities.
The system integrates specialized agents for parameter generation, mechanics-based calculation, target-interval evaluation, deviation diagnosis, and rule-based design modification, according to arxiv.org. Researchers evaluated the framework on sixty RC barrier designs across three testing levels (TL-3, TL-4, and TL-5) using AASHTO LRFD Bridge Design Specifications, 10th Edition (2024).
Three DeepSeek models with different parameter scales were tested: DS-8B, DS-32B, and DS-671B. The lightweight MAF-DS-8B model achieved the 98.3% compliance rate, while the standalone DS-32B model reached only 11.7%, demonstrating that “multi-agent architectures can improve the reliability, interpretability, and accessibility of AI-assisted engineering design for practical applications,” according to the researchers.
The framework specifically targets the safety-critical task of designing reinforced concrete highway barriers, which “requires strict compliance with regulatory provisions,” the paper states.