Conceptual engineering system design faces challenges from traditional methods and emerging AI tools to fully address its inherently complex, dynamic, and creativity-driven demands. iDesignGPT is a framework that integrates large language models with established design methodologies to enable dynamic multi-agent collaboration for problem refinement, information gathering, design space exploration, and evaluation. By incorporating design metrics such as coverage, diversity, and novelty, iDesignGPT provides quantitative insights for early-stage conceptual design. Performance evaluations across six public design challenges show that iDesignGPT achieves competitive novelty and consistently higher originality and modularity than GPT-4o zero-shot, GPT-4o chain-of-thought and Deepseek-r1, based on metrics and expert assessments. Two controlled user studies show positive reception across profiles and, for novice designers, lower mental demand than human-only design and clearer design flow with iDesignGPT. These results establish iDesignGPT as a practical framework for integrating language-model agents with established engineering design methods, enabling metrics-driven support for conceptual design by both expert and novice designers.
Accurately predicting the material loss rate is critical for MRP-driven delivery and inventory control. Inaccurate and delayed data acquisition leads to imprecise master data and poor knowledge of how machines are used in small- and medium-sized manufacturing enterprises (SMEs), but dynamic production environments and data latency are persistent challenges. This paper proposes a two-stage dynamic prediction framework integrating Time Decay Weighted Average (TDWA) and AHP-Gray Markov correction. TDWA establishes a baseline loss rate prioritizing recent process data, while the Analytic Hierarchy Process (AHP) quantifies eleven heterogeneous factors affecting material consumption. A Gray model predicts loss-rate overshoot, refined by Markov chain residual correction. Empirical validation using 42 months of production data from a hardware–sanitary ware manufacturer demonstrates a MAPE of 1.97%, outperforming standalone by 38% and the three-year average by 66.5%. The method effectively balances delivery assurance and inventory costs, maintaining zero stockouts while limiting excess inventory to 1.7%. Unlike ARIMA, XGBoost, or LSTM, which suffer from overfitting or performance degradation with eight or fewer data periods, the proposed model maintains high accuracy and robustness with extremely small samples. Offering a low-threshold “small-sample master data governance” paradigm, this approach may be applied to aerospace spare parts support, cold chain logistics, and customized manufacturing, providing a viable pathway for digital transformation in resource-constrained environments.
The shipbuilding industry plays a pivotal role in national strategic security and economic development, and one critical challenge is the pipeline layout design problem. It predominantly relies on designers’ subjective experience, and it is marked by a lack of efficient knowledge sharing and the absence of smart pipe routing algorithms. This paper proposes an agile design system, which integrates ship pipeline design knowledge management, semi-automatic design that involves frequent interaction with human designers, and automatic rule checking. The framework is refined by digital twin concepts, facilitating close collaboration between physical and digital systems. The paradigm shift holds the potential to substantially enhance the efficiency of ship pipeline layout design, while concurrently reducing the reliance on manual labor.