Scholarlink

Abstract <p>Large language models (LLMs) and retrieval-augmented generation (RAG) are transforming how knowledge is represented, retrieved, and reasoned upon in materials science. Traditional discovery workflows depend on time- and cost-intensive simulations and experiments to navigate vast compositional and structural spaces. By integrating language understanding with domain-specific retrieval and tools, RAG-enabled LLMs can accelerate this process, automating literature mining, proposing crystal structures, analyzing defects, and generating hypotheses grounded in both data and physics. This article provides a systematic and theory-forward review of LLMs, RAG, and related agentic architectures for crystalline materials discovery and characterization. Using a PRISMA-inspired protocol, we survey work from 2019&#xa0;to&#xa0;2025 across six application areas: crystal structure prediction, defect analysis, materials search and optimization, literature mining, database integration, and multimodal reasoning. Building on this background, we present three complementary theoretical lenses: (1) a materials science lens that frames discovery as search and optimization over high-dimensional energy and structure landscapes; (2) a decision-theoretic lens that models LLM+tools+physics systems as query-act-observe processes (POMDP-style) over materials design spaces; and (3) a multi-agent lens that interprets emerging architectures (e.g., scientist, planner, critic, and tool-using agents) as hierarchical and modular policies coordinated through shared memories or materials knowledge graphs. We further present a theory-driven view of evaluation and hallucination based on risk-coverage tradeoffs, selective prediction, and physics- and database-aware consistency checks. An integrated comparison table contrasts LLM-based approaches with non-LLM state-of-the-art not only in terms of performance, but also in terms of their inductive biases and priors over the design space. Finally, we outline human-in-the-loop multi-agent designs and a set of theory-forward open questions on optimal use of retrieval, tool policies, and LLM-induced priors over energy landscapes. Together, these elements position LLM+RAG systems as analyzable, tunable scientific instruments rather than opaque heuristics, and chart a path toward trustworthy AI collaborators in crystalline materials discovery.</p> Graphical abstract <p>This visual abstract illustrates how large language model (LLM)-based AI agents can sit at the center of end-to-end materials science workflows. On the left, materials science tasks such as materials discovery and materials characterization define the high-level scientific goals. In the center, these tasks are linked to a large pool of unstructured and structured scientific knowledge, literature, curated databases, and multimodal inputs (text, structures, images, spectra). AI/LLM agents interact with this knowledge base in both single-agent and multi-agent configurations: individual agents query and reason over the data, while multi-agent networks coordinate specialized roles (e.g., retrieval, planning, simulation control, and critique). Through this closed loop, the agents transform raw scientific information into accelerated discovery and insight generation on the right, producing new crystal structures, property predictions, and literature-grounded summaries that directly support experimental design and theoretical understanding in materials research. </p> Impact statement <p>Large language models and retrieval-augmented generation are rapidly emerging as tools for scientific discovery, yet their role in materials science remains conceptually fragmented. This work provides a systematic and theory-driven synthesis of how LLM-based systems can support crystalline materials discovery by integrating language reasoning, domain retrieval, and physics-based tools. By framing these systems through energy-landscape optimization, decision-theoretic reasoning, and multi-agent scientific workflows, this review establishes a conceptual foundation for designing reliable AI-assisted materials discovery platforms.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLMs, RAG systems, and agents in crystalline materials discovery and characterization: A systematic review

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Joseph Oche Agada,
  • Arpan Biswas

摘要

Abstract

Large language models (LLMs) and retrieval-augmented generation (RAG) are transforming how knowledge is represented, retrieved, and reasoned upon in materials science. Traditional discovery workflows depend on time- and cost-intensive simulations and experiments to navigate vast compositional and structural spaces. By integrating language understanding with domain-specific retrieval and tools, RAG-enabled LLMs can accelerate this process, automating literature mining, proposing crystal structures, analyzing defects, and generating hypotheses grounded in both data and physics. This article provides a systematic and theory-forward review of LLMs, RAG, and related agentic architectures for crystalline materials discovery and characterization. Using a PRISMA-inspired protocol, we survey work from 2019 to 2025 across six application areas: crystal structure prediction, defect analysis, materials search and optimization, literature mining, database integration, and multimodal reasoning. Building on this background, we present three complementary theoretical lenses: (1) a materials science lens that frames discovery as search and optimization over high-dimensional energy and structure landscapes; (2) a decision-theoretic lens that models LLM+tools+physics systems as query-act-observe processes (POMDP-style) over materials design spaces; and (3) a multi-agent lens that interprets emerging architectures (e.g., scientist, planner, critic, and tool-using agents) as hierarchical and modular policies coordinated through shared memories or materials knowledge graphs. We further present a theory-driven view of evaluation and hallucination based on risk-coverage tradeoffs, selective prediction, and physics- and database-aware consistency checks. An integrated comparison table contrasts LLM-based approaches with non-LLM state-of-the-art not only in terms of performance, but also in terms of their inductive biases and priors over the design space. Finally, we outline human-in-the-loop multi-agent designs and a set of theory-forward open questions on optimal use of retrieval, tool policies, and LLM-induced priors over energy landscapes. Together, these elements position LLM+RAG systems as analyzable, tunable scientific instruments rather than opaque heuristics, and chart a path toward trustworthy AI collaborators in crystalline materials discovery.

Graphical abstract

This visual abstract illustrates how large language model (LLM)-based AI agents can sit at the center of end-to-end materials science workflows. On the left, materials science tasks such as materials discovery and materials characterization define the high-level scientific goals. In the center, these tasks are linked to a large pool of unstructured and structured scientific knowledge, literature, curated databases, and multimodal inputs (text, structures, images, spectra). AI/LLM agents interact with this knowledge base in both single-agent and multi-agent configurations: individual agents query and reason over the data, while multi-agent networks coordinate specialized roles (e.g., retrieval, planning, simulation control, and critique). Through this closed loop, the agents transform raw scientific information into accelerated discovery and insight generation on the right, producing new crystal structures, property predictions, and literature-grounded summaries that directly support experimental design and theoretical understanding in materials research.

Impact statement

Large language models and retrieval-augmented generation are rapidly emerging as tools for scientific discovery, yet their role in materials science remains conceptually fragmented. This work provides a systematic and theory-driven synthesis of how LLM-based systems can support crystalline materials discovery by integrating language reasoning, domain retrieval, and physics-based tools. By framing these systems through energy-landscape optimization, decision-theoretic reasoning, and multi-agent scientific workflows, this review establishes a conceptual foundation for designing reliable AI-assisted materials discovery platforms.