Automating Structural Analysis Across Multiple Software Platforms Using Large Language Models

A team of researchers from Lehigh University and other institutions has developed a novel AI system that automates the creation of structural analysis models across multiple, distinct software platforms. The work, detailed in a new arXiv paper, addresses a critical gap in previous LLM applications for engineering, which were typically limited to a single software environment. In practice, structural engineers often need to use different Finite Element Analysis (FEA) tools like ETABS, SAP2000, and OpenSees depending on project specifications or company standards. This new system breaks that limitation by employing a sophisticated two-stage, multi-agent architecture.

In the first stage, a cohort of specialized AI agents collaboratively interprets a user's natural language or schematic input. They perform structured reasoning to infer all necessary modeling parameters—including geometry, materials, boundary conditions, and loads—and compile this information into a unified, software-agnostic JSON representation. In the second stage, the system deploys parallel code translation agents. Each agent is an LLM specifically prompted with the syntax and modeling workflows of a target platform (ETABS, SAP2000, or OpenSees) and converts the central JSON file into a ready-to-run script.

The researchers rigorously evaluated the system using 20 representative frame analysis problems. Over ten repeated trials for each case, the multi-agent LLM demonstrated consistently reliable performance, achieving accuracy rates exceeding 90% across all three software platforms. This high level of reliability is crucial for real-world engineering adoption, where errors can be costly. The architecture's design not only boosts accuracy but also offers inherent flexibility, allowing new software targets to be added by simply introducing a new translation agent with the appropriate prompts, paving the way for broader industry application.