2024 AIChE Annual Meeting

(664d) Prompt Engineering-Directed in-Context Learning of an LLM with Integrated Automated Simulation for Generative Chemical Process Design

Authors

Shin, D., Myongji University
Chemical process design involves the optimal design of large-scale chemical processes among various candidates and requires considerable time and effort to navigate extensive search space constrained by cost-effectiveness, safety, environmental friendliness, and market conditions. Various AI-based methodologies have been proposed, such as finding the optimal process through reinforcement learning or outputting the best process through greedy search and simulations. However, these methodologies exhibit two critical drawbacks: low versatility, making it challenging to apply the methodologies to diverse processes, and low explainability, making it hard to explain how the optimal process was chosen.

This study proposes a methodology to overcome the low versatility and explainability by incorporating a Large Language Model (LLM), Generative Pre-trained Transformer (GPT), into the chemical process design process. To conduct process diagram input and analysis through a language model, we use the Detailed Simplified Flowsheet Input Line Entry System format that includes details of flows and equipment. Furthermore, we apply various prompt engineering techniques to direct the in-context learning of the LLM, integrated with automatically generated process simulations, to transform the general solver GPT into an intelligent agent specialized in chemical process design. Prompt chaining is applied to divide and solve the complex problem of chemical process design sequentially, breaking down the design and improvement process. Techniques such as retrieval augmented generation (RAG) for diverse processes application and In-Context Learning (ICL), automatic reasoning and tool-use (ART) for simulation integration, and directional stimulus prompting to focus on important information are applied to construct prompts specialized for chemical process design and to proceed with the enhanced design and improvement processes.

By applying the proposed system to various case studies, we conducted process design and improvement, demonstrating process improvements through various methods such as reactor addition and recycling. The improved results showed that diverse generic process improvements are possible. Moreover, the improved process was analyzed through the language model's inherent explanation capability and the additional integration with the simulator DWSIM, showing improved explainability. Ultimately, the study presents a new methodology for automating and supporting process design based on intelligent generation and validation of alternative processes guided by LLM-based design agents.