Sang-Gook Kim

With the rapid rise of Large Language Models capable of representing intricate multi-modal data, decision-makers facing contingencies can be contextually supported with real-time data, dynamic simulations and actionable knowledge insights. This can address critical challenges in production planning and control, particularly in managing stochastic, unforeseen or uninformed changes, i.e. during contingencies. A framework of localized LLM agents for a specific manufacturing environment and system is used to collect past and current decision reasonings, sensory and simulation data to propose contingency management reactions. This concept is implemented within a small-scale semiconductor manufacturing system, demonstrating a unique opportunity for the successful digital transformation of production planning and control.

Abstract The increasing uncertainty in modern manufacturing systems presents challenges in managing disruptions. Unlike traditional linear production systems, matrix production enables modular flexibility, allowing for real-time machine rescheduling and dynamic material transport via Automated Guided Vehicles (AGVs). However, this added flexibility also introduces substantial complexity in decision-making, particularly under disruption scenarios, as it necessitates the simultaneous adjustment of both AGV dispatching and machine scheduling. Traditional optimization- and rule-based approaches to disruption management often suffer from extended decision-making times, suboptimal performance, and, critically, the limited integration of human expertise. To address these limitations, we propose a Generative AI (GenAI)-driven decision support framework that integrates Large Language Models (LLMs) within matrix production systems to enhance manufacturing resilience. Our framework leverages a fine-tuned GenAI model trained on structured knowledge bases derived from digital twin-enabled what-if disruption scenarios and expert knowledge. By integrating human-guided decision refinement, the framework provides real-time, human-in-the-loop, and explainable decision support aligned with operational goals and human preferences. To validate the approach, we implemented the framework in a semiconductor production use case and assessed the system resilience via key metrics such as utilization, downtime, and throughput. Compared to the best baseline static rule, our GenAI-based framework improved throughput by 11.81% and machine utilization by 11.49% in disruption scenarios. The results show that task-specific recommendations generated by the GenAI model effectively assist workers in selecting optimal response decisions during real-time disruptions.