Large language models in materials science: mechanistic understanding, structure–property reasoning and future opportunities

Materials science is governed by complex, high-dimensional relationships among composition, processing, structure and properties, creating vast design spaces that make materials development inefficient. Large language models (LLMs) are emerging as powerful tools for navigating this complexity through scientific knowledge processing and reasoning. However, applying LLMs to materials science requires more than transferring general artificial intelligence methods to a new domain; it demands evaluating whether these models can support physically grounded, mechanistically meaningful and materials-specific reasoning. Here, we first introduce the fundamental principles of LLMs and optimization strategies for downstream scientific tasks. We then discuss materials language processing as a foundation for structuring materials knowledge and enabling downstream tasks. Building on this basis, we review the emerging roles of LLMs in mechanistic understanding and structure–property relationships, including property prediction, knowledge discovery and inverse design. We further summarize domain-specific materials LLMs, LLM-based agents, autonomous laboratories and benchmarks in materials science. Finally, we discuss current limitations and future opportunities for mechanism-aware domain-specialized models, multimodal scientific learning, real-world evaluation and integrated AI systems. Ultimately, the future impact of LLMs in materials science will depend on whether they can evolve toward physically grounded reasoning systems for scientific materials discovery.