Knowledge-guided large language model for material science
Guanjie Wang, Jingjing Hu, Jian Zhou, Sen Liu, Qi Li, Zhimei Sun
Abstract
With ChatGPT starting a storm of transformative applications worldwide, the advent of large language models (LLMs) has revolutionized the paradigm of scientific research, shifting from data-driven methods to AI-driven science. While LLMs have demonstrated significant promise in many fields of science, the development of material knowledge-guided, domain-specific LLMs remains challenges. In this review, the key milestones of LLMs are discussed, and guidelines for building LLMs are provided, including determining objectives, designing model architectures, data curation, and establishing training and evaluation frameworks. Furthermore, methodologies for creating domain-specific models through fine-tuning, retrieval-augmented generation, prompt engineering, and AI agents are explored. Additionally, the applications of LLMs in materials science are investigated, ranging from structured information extraction and property prediction to autonomous laboratories and robotics. Finally, challenges such as resource demands, dataset quality, benchmarking, hallucination mitigation, and AI safety are reported alongside emerging opportunities, positioning LLMs as a pivotal tool in advancing materials discovery and innovation.