Abstract:
Large language models are igniting a quiet revolution in how we practice materials science. What began as tools for language and code are rapidly becoming engines for scientific discovery, capable not only of reading our literature, but of designing our materials. In this talk, I trace a new end-to-end paradigm that runs from unstructured text to engineered crystal structures. (1) I begin with KnowMat, our agentic, LLM-driven framework for transforming the materials literature into structured, machine-readable data. KnowMat converts PDFs, tables, and narrative text into validated JSON schemas, enabling automated database construction, large-scale literature mining, and ML-ready datasets with built-in quality control. This shifts data curation from a manual bottleneck into an automated, scalable scientific instrument. (2) I then move from understanding to creation. I introduce CrysText, a framework that uses LLMs to generate full crystallographic information files (CIFs) directly from natural-language prompts. Rather than treating crystal generation as a purely numerical problem, CrysText treats it as a language problem, allowing composition, symmetry, and even thermodynamic stability to be specified in text. With reinforcement learning layered on top, these models learn not just to speak crystallography, but to obey its physics. Together, KnowMat and CrysText define a new closed loop for materials discovery: literature → structured data → generative design → candidate materials. LLMs are no longer just assistants to materials scientists; they are becoming co-designers.