Enterprise AI Analysis

The novel multi-agent architecture decomposes the structural modeling workflow into four coordinated modules: analysis and planning, geometry assembly, load integration, and code translation. It features problem analysis and construction planning agents, parallel node and element agents, a load assignment agent, and code translation agents. Checkpoints are embedded for consistency checks and regeneration, mitigating error propagation. The architecture uses two lightweight LLM backbones: GPT-OSS 120B for complex reasoning and Llama-3.3 70B Instruct Turbo for information translation, leveraging their complementary strengths.

The proposed multi-agent architecture was evaluated on a benchmark of 20 frame problems, demonstrating strong reliability and robustness with 99% average accuracy (100% in 18 cases, 90% in two). It significantly outperforms sequential multi-agent architectures (91% accuracy) and general-purpose LLMs (e.g., Gemini 2.5 Pro at 37%, GPT 40 at 0%). The architecture also reduces inference runtime from 269.2-949.0 seconds to 75.4-194.2 seconds, an 85% reduction for complex cases, and shows excellent scalability to larger structural systems (7-bay, 10-bay frames) with high accuracy rates (90%-100%).

Despite its robust performance, the architecture has limitations. Prompt templates, while effective for selected LLMs, show sensitivity across different models, highlighting the need for automated prompt optimization. The current system is restricted to rectangular frames and lacks support for components like diagonal bracing or cantilevers. Additionally, the analysis scope is limited to linear elastic behavior under static loads, omitting dynamic effects (wind, seismic) or nonlinear analysis. Future work will focus on enhancing generalizability to complex typologies, supporting advanced analysis capabilities, and improving prompt robustness.