Closed-Loop CNT Growth Using Retrieval Augmented Generation and Human Feedback

Carbon Nanotube (CNT) synthesis is a complex process requiring precise control of input parameters in scanning electron microscopy (SEM) experiments by humans to achieve desired material properties. Use of artificial intelligence (AI) in such a human-machine system can foster a closedloop design to accelerate material discovery via integration of theory, computation, and experimentation. In this paper, we adopt emerging generative AI advances i.e., Large Language Models (LLMs) to guide human-machine interactions in CNT growth experiments and investigate an intelligent architecture viz., “ARES-Vidura” featuring ML models and human feedback to analyze CNT domain-specific datasets. ARESVidura features a knowledge base and Retrieval-Augmented Generation (RAG) to enhance user queries and responses with LLMs, simulation tools for CNT synthesis, and regression models for guiding novice/experienced researchers in predicting synthesis parameters using synthetic and real data training that alleviate data scarcity by mimicking experimental conditions. Our experimental results show that our approach delivers over 89% better performance in terms of response relevance for a novice user, as measured by the cosine similarity, outperforming public LLMs (e.g., GPT), with the Llama-2-7B model showing the best results, while the regression component achieves less than 2% mean square error in predicting CNT growth rate.