GPT Prompt Engineering for a Large Language Model-Based Process Improvement Generation System

Abstract

Process design improvements require extensive knowledge, considerable time, and huge human resources due to the complexity of chemical processes and their diverse objective functions. However, machine learning-based approaches using vast accumulated data are limited in low versatility, applicable only to specific processes, and unable to understand the basis of model decisions. This study proposes the GPT-based Improved Process Hybrid Transformer (GIPHT), a process design improvement generation system utilizing Large Language Model (LLM). LLMs, being natural language-based, allow for understanding the basis of model decisions without need of explainable AI analysis. GIPHT is composed of multi-agent to enhance versatility and performance for diverse chemical processes. We also propose the Detailed Simplified Flowsheet Input Line Entry System format to express process diagrams in natural language, including enhanced information about process conditions. A structured prompt system is employed and validated in the LLM domain through prompt engineering. GIPHT searches and extracts data based on its proposed improvement methodology, providing explanations for the decision-making process and the basis, overcoming limitations of the traditional black-box AI models. It offers directional ideas to design engineers in the early stages of process design and would be used for training of process engineers, supporting improvement of outdated processes and transformation into more environmentally friendly processes.