Page 111 - Read Online

Li et al. J. Mater. Inf. 2026, 6, 10 Page 3 of 31

characterization to enhance research efficiency, and analyzing complex experimental data to uncover
fundamental relationships. These applications not only improve synthesis efficiency but also reduce
experimental costs, making materials research more precise and controllable. More importantly, agents have
driven the development of SDLs, which effectively integrate theoretical design in the computational domain
with experimental validation in the experimental domain, marking the advent of a new era of automated and
intelligent materials research [17,18] .

This review aims to systematically analyze the recent progress and development trends in the application of
agents in material science. Following the logical flow of materials research and development, the article is
structured around three core phases: materials design, materials synthesis and characterization, and
autonomous laboratories. The first section focuses on the role of agents in materials design, highlighting
their applications in literature data extraction, material knowledge question answering, scientific hypothesis
generation, material structure design, and theoretical property calculation. The second section discusses the
application of agents across the entire process of materials synthesis and characterization, including material
synthesis processes exploration, autonomous material characterization, experimental data analysis, and
science theory discovery. The third section presents specific examples of agent-driven SDLs and reviews their
developmental process. Through this systematic review, the article aims to provide a comprehensive
perspective on AI agent technologies for material science researchers, facilitate the integration of AI and
material science, and accelerate the innovation and application of new materials.

AGENTIC MATERIALS DESIGN
Materials design is a process of pursuing innovation and breakthroughs from vast information and complex
rules. Its key challenge lies in the efficient identification of materials with target properties from an expansive
materials space. In the traditional paradigm, researchers had to manually search literature, analyze data,
propose hypotheses, and perform computational verification. This process is not only inefficient but also
limited by human cognitive constraints. The introduction of AI agents is fundamentally transforming this
paradigm. By deeply integrating the cognitive capabilities of LLMs with the computational power of
professional tools, they form an autonomous design system spanning from data extraction to theoretical
design. This section will systematically analyze how AI agents empower key stages of materials design and
elaborate on the implementation pathways and technological breakthroughs across five critical tasks.

Data extraction
Scientific literature is the largest repository of materials knowledge, and the structured processing of
literature data forms the foundation of intelligent applications. However, its unstructured form renders
manual extraction inefficient and susceptible to omission of key information or errors. Agents based on
LLMs, through their powerful natural language understanding, multimodal processing, and multi-tool
coordination capabilities, effectively overcome the challenges of scattered, heterogeneous, and large-scale
literature data. They significantly improve the efficiency and accuracy of data extraction, enabling efficient
and precise knowledge extraction from scientific literature [19-21] .

These systems typically employ multi-agent systems (MAS), where role function definition and task
decomposition are used to enhance the comprehensiveness and accuracy of data extraction. For example,
ChemMiner incorporates three specialized agents for text analysis, multimodal analysis, and synthesis
analysis, to process complex coreference mapping relationships, non-textual information, such as figures and
tables, and final data generation, respectively. This enables end-to-end extraction of reaction information
from chemical literature, achieving performance close to that of human chemists while significantly reducing
processing time [Figure 1]. Similarly, nanoMINER uses a ReAct (Reasoning and Acting) agent to
[22]
coordinate multiple specialized agents, together with visual tools such as You-Only-Look-Once (YOLO) for