Multimodal Large Language Models for Cross-Domain Knowledge Transfer and Integration in Intelligent Engineering Systems

Abstract

Multimode Wide Language Model (MLLM) is changing the intelligent engineering system. They allow unified reasoning of different data streams in a complex network physical environment. Based on system engineering theory, network physical system (CPS) principle and information theory, this review analyzes how MLLM can help interdisciplinary knowledge transfer. They use attention-based alignment, adjustment instructions and mixed integration with digital twins, reinforcement learning, computer vision and edge computing. Practical cases from semiconductor manufacturing, independent logistics, smart grid and infrastructure inspection show the progress of predictive maintenance, dynamic scheduling and real-time anomaly detection. Although MLLM reduces the separation between domains and accelerates adaptability, it still has important limitations. For example, there is a risk of illusion in the safety-critical control loop, and the computational load is sometimes too heavy for edge computing, so it is difficult to understand the physical world well. The initial contribution includes a multimodal information bottleneck framework for efficient transfer and elasticity index, and for measuring redundant knowledge paths in CPS. The review concludes that the mixed design of neural symbols and physical information, coupled with strict verification, is essential for reliable deployment. Future progress will require interdisciplinary progress in artificial intelligence, control theory and dependency engineering.