Application of literature datasets and knowledge transfer to accelerate material design

Traditional material development faces challenges such as long experimental cycles and high costs. While high-throughput computing can quickly screen candidate materials, the predictive results often exhibit significant discrepancies from experimental data. By leveraging large language models (LLMs) and Retrieval-Augmented Generation (RAG) technology, a high-quality MLCC material dataset was constructed by extracting data from a vast array of scientific literature. Subsequently, the pre-trained BERT model was fine-tuned using Low-Rank Adaptation (LoRA) technology, successfully developing a machine learning model with high predictive accuracy (R2 value of 0.95). This model can accurately predict the dielectric properties of complex perovskite materials, providing a new approach for efficient research and development in the field of material science. This method not only overcomes the limitations of traditional computational methods but also lays the groundwork for future material science research.