Digital Twin Builder: A Multi-Agent LLM System for Automated Industrial Digital Twin Development

Maria Dziuba, Elizaveta Egorova, Nikolay Vasilev, Ilya Novitskiy, Vladislav Tuchin, Valeria Efimova

Keywords: Computer Science, Large Language Models (LLMs), Multi-agent Systems, Digital Twins

Abstract:

Multi-agent LLM systems are increasingly applied to complex code generation tasks, including industrial software, but most existing systems are narrow, lack safety guaranties, and cannot be reused across production scenarios.
We present Digital Twin Builder, an end-to-end multi-agent system that builds industrial digital twins (DTs) from natural-language descriptions of manufacturing processes. An orchestrator coordinates specialized agents for user interaction, schema design, code generation, and a multi-modal complex anomaly detector that fuses sensor and camera streams. Safety is enforced via schema-grounded synthesis with structural and functional test gates at evaluation time, plus bounded fail-closed orchestration at runtime.
On a steel-casting/cement-plant benchmark, our pipeline attains 100\% Structural Test Rate, $5.5\times$ over a single-LLM ablation, within 10 J-LLM points of Claude 4.6 Sonnet; a blind study over 118 scenarios confirms decomposition as the main driver of functional completeness.

Track: Short Paper (4 pages)

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 269