Finite element analysis is widely used to predict the stress and deformation behavior of structures, but constructing analysis models and assigning boundary conditions require specialized expertise, limiting accessibility for non-experts. Recently, AI agents based on large language models (LLMs) have been actively studied. However, commercial LLMs cannot access confidential corporate data. They also carry the risk of hallucination, where incorrect values may be reflected in the analysis results. In addition, security concerns often prevent their use with proprietary design data. To address these limitations, this study employed open-source LLM that can be operated within a closed network, to automatically extract design conditions from natural language input, and introduced a parsing and validation procedure to prevent hallucination-induced errors. The validated data were reflected in the analysis model by precisely substituting only pre-defined tagged variables within the pre-processing python script, thereby modifying the model without altering the remaining script structure. The proposed system was applied to a case study involving the design modification of robot arms in a multi-joint robot used for ship-building industry; the cross-sectional shape, length, and material of the robot arms specified through natural language input were accurately reflected in the generated script, while unintended parts remained unchanged. These results demonstrate that users without expertise in finite element analysis can modify analysis models and obtain results solely through a conversational interface.

Abstract
Abstract
1. 서 론
2. Proposed model
    2.1 Overall Workflow
    2.2 Local LLM-Based Interface
    2.3 Automatic modification of the script
3. Case study
    3.1 Application to multi-joint robot
    3.2 Result
4. Discussion
5. Conclusion
Acknowledgements
References

• 표창민(Senior Researcher, Korea institute of industrial technology) | Changmin Pyo Corresponding author