摘要
Process simulation software such as Aspen Plus has become indispensable in chemical engineering design, yet the translation of engineering intent into software operations remains a significant barrier for practitioners. This study presents a Large Language Model (LLM)-powered Copilot framework that integrates large language models (Claude Sonnet 4.5) with the Model Context Protocol (MCP) to enable human-AI collaborative process design. Unlike fully automated approaches, the proposed Copilot paradigm positions engineers as decision-makers while delegating routine software operations to AI assistants. The framework comprises 42 modular tools organized into 7 functional categories, covering the complete simulation workflow from model initialization to result analysis. A hierarchical Skills knowledge system was developed to guide AI operations and reduce hallucination errors. The effectiveness of the framework was validated through three case studies: (1) water-ethanol binary distillation achieving consistent semantic accuracy across six independent runs with +/- 0.1% purity deviation after fine-tuning; (2) pressure-swing distillation where the Copilot proactively identified thermodynamic limitations of the azeotrope system and autonomously optimized to 97.86 wt % ethanol; and (3) literature-based isopropyl alcohol (IPA) extractive distillation reconstruction with full structure fidelity and <1% energy duty error. The results demonstrated that the Copilot achieved high structural accuracy in model construction while substantially reducing model construction time compared to manual operations. The human-AI collaborative approach maintained engineering oversight while lowering the expertise barrier for process simulation. These findings highlight the potential of LLM-powered Copilots to transform chemical process design by enabling engineers to focus on design objectives rather than software mechanics.