Gen_AI_DES: Automated Generation and Constrained Optimization of Manufacturing Simulations Using Large Language Models and Knowledge Graphs : LLM-Driven Simulation with Knowledge Graph Verification

Sammanfattning

Manufacturing companies often use computer-based simulation models to evaluate improvements in production systems before making physical changes. These simulations can help reduce cost and risk, but building them accurately from production data or event logs is typically time-consuming and requires specialized modelling expertise.Recent advances in artificial intelligence, in particular Large Language Models, make it possible to generate simulation models automatically from a description written in plain language. However, such automatically generated models are not always reliable, and they often require manual correction to ensure that they are structurally correct and consistent with the underlying production system. This limits their practical use in industrial environments.This thesis investigates how to improve the reliability of automatically generated simulation models by introducing structured verification methods based on formal representations of system knowledge. Two complementary approaches are presented.The first approach combines automatic model generation with a structured validation step. A simulation model is created from production event data and then checked against a formal description of the system structure. If inconsistencies are detected, detailed error messages are used to automatically guide the correction of the model. This process is integrated into a user-friendly interface that does not require programming skills. The method is tested on an automotive production line and shows that the system can identify structural errors and produce a valid simulation model after one correction cycle without human intervention. The results also show good agreement with a commercial simulation tool in terms of predicted production throughput.The second approach integrates structural knowledge earlier in the modelling and optimization process. Here, information about system parameters is organised into a structured representation that clearly defines valid model configurations. This representation is used to guide both the construction of the simulation model and the search for improved operating settings. The method is tested on a carbon-fibre production line and produces a detailed structured model of the system. It also identifies improved operating conditions that increase production output while reducing energy use, with results closely matching a validated reference simulation tool.Together, these two approaches demonstrate that introducing structured, verifiable representations of system knowledge can significantly improve the reliability and usefulness of automatically generated manufacturing simulation models. This provides a practical pathway toward more automated and trustworthy simulation-based decision support in industrial settings.