The convergence of digital twins, manufacturing execution systems, and generative artificial intelligence has produced a new epistemic and technological configuration for contemporary production systems. In highly complex manufacturing environments characterized by volatile demand, high product variety, and stringent efficiency requirements, the traditional rule-based or heuristics-driven optimization of manufacturing execution systems has increasingly demonstrated its structural limitations. This article develops a comprehensive theoretical and methodological framework for understanding how large language model–driven generative artificial intelligence can be integrated with digital twin–enabled cyber-physical production systems to provide dynamic, context-aware, and lean-compatible optimization of manufacturing execution system configurations. Building on the recent conceptual and empirical insights offered by Chowdhury, Pagidoju, and Lingamgunta in their analysis of generative AI for MES optimization, this study situates LLM-based recommendation engines within the broader intellectual traditions of lean manufacturing, operations research, discrete-event simulation, and digital twin theory.

The article advances the argument that manufacturing execution systems should no longer be conceptualized merely as transactional information platforms but as adaptive cognitive infrastructures embedded in cyber-physical ecosystems. Through the integration of digital twin architectures standardized under ISO 23247 and data- and knowledge-driven modeling frameworks, MES platforms can serve as real-time operational mirrors of the shop floor. Generative AI, particularly large language models, introduces a fundamentally new layer of interpretive and configurational intelligence that allows these mirrors to not only reflect but also reason about production states, constraints, and improvement opportunities. Drawing on lean management theory, including Toyota Kata and Gemba Kaizen, the article argues that LLM-driven MES optimization can support continuous improvement by translating tacit operational knowledge into executable system configurations.

Methodologically, the article adopts a multi-layered conceptual synthesis grounded in simulation theory, cyber-physical systems design, and decision support architectures. It critically examines how NP-complete scheduling problems, traditionally addressed through algorithmic approximations and simulation-based heuristics, can be reframed through generative reasoning that dynamically explores configuration spaces. The results section develops a theoretically grounded model of how LLM-based recommendation engines, when connected to digital twins of manufacturing cells, can improve responsiveness, reduce configuration inertia, and enhance alignment between strategic objectives and shop-floor realities. The discussion situates these findings within broader debates on lean and Industry 4.0 integration, energy-aware production, supply chain coordination, and human–machine collaboration.

By synthesizing insights from manufacturing science, information systems, and artificial intelligence, this article contributes a novel conceptual architecture for intelligent MES optimization. It demonstrates that the future of manufacturing control lies not in replacing human expertise but in augmenting it through generative, explainable, and context-sensitive digital companions that operate within digital twin ecosystems.