Large Language Models (LLMs) show promise in accelerating materials science research but currently lack the depth and accuracy needed for practical applications. This paper identifies key limitations of LLMs in materials science, including difficulties in reasoning over complex, interconnected knowledge and handling domain-specific language and notation. To address these challenges, the authors propose a framework for developing Materials Science LLMs (MatSci-LLMs) that integrate domain knowledge and hypothesis generation with testing. The development of effective MatSci-LLMs requires high-quality, multi-modal datasets derived from scientific literature, which capture diverse materials knowledge. The paper outlines a roadmap for applying MatSci-LLMs in real-world materials discovery through three key steps: automated knowledge base generation, automated in-silico material design, and integrated self-driving materials laboratories. The paper also highlights current failures of LLMs in materials science, such as poor performance in numerical reasoning and crystal structure interpretation. These failures underscore the need for more domain-specific training and improved reasoning capabilities. The paper emphasizes the importance of multi-modal information extraction, including text, tables, figures, and videos, to fully capture materials science knowledge. Finally, the paper discusses the broader impact of MatSci-LLMs on materials discovery, including potential benefits for sustainability, energy, and healthcare, as well as the need for ethical considerations in their application.Large Language Models (LLMs) show promise in accelerating materials science research but currently lack the depth and accuracy needed for practical applications. This paper identifies key limitations of LLMs in materials science, including difficulties in reasoning over complex, interconnected knowledge and handling domain-specific language and notation. To address these challenges, the authors propose a framework for developing Materials Science LLMs (MatSci-LLMs) that integrate domain knowledge and hypothesis generation with testing. The development of effective MatSci-LLMs requires high-quality, multi-modal datasets derived from scientific literature, which capture diverse materials knowledge. The paper outlines a roadmap for applying MatSci-LLMs in real-world materials discovery through three key steps: automated knowledge base generation, automated in-silico material design, and integrated self-driving materials laboratories. The paper also highlights current failures of LLMs in materials science, such as poor performance in numerical reasoning and crystal structure interpretation. These failures underscore the need for more domain-specific training and improved reasoning capabilities. The paper emphasizes the importance of multi-modal information extraction, including text, tables, figures, and videos, to fully capture materials science knowledge. Finally, the paper discusses the broader impact of MatSci-LLMs on materials discovery, including potential benefits for sustainability, energy, and healthcare, as well as the need for ethical considerations in their application.