Materials science presents a highly challenging domain for information retrieval systems due to its dense numerical data, chemical content, and special terms. Traditional Retrieval-Augmented Generation (RAG) systems struggle to effectively process the complex relationships between material properties, structural parameters, and characterization data. A dynamic hybrid retrieval framework is proposed that combines keyword extraction rate with query perplexity metrics. By integrating chain of thought reasoning and self-consistency techniques in the generation phase, our system effectively navigates the inherent numerical and symbolic complexity of materials science discourse. Extensive experiments on comprehensive materials datasets demonstrate significant improvements in both retrieval quality and response accuracy across various model architectures.