Page 110 - Read Online

Page 2 of 31 Li et al. J. Mater. Inf. 2026, 6, 10

INTRODUCTION
Material science, as a pivotal discipline underpinning national strategic demands such as energy transition,
biomedicine, and advanced manufacturing, directly determines the progress of key technologies and
industrial competitiveness. Its central challenge lies in how to efficiently discover, design, and synthesize
novel materials with targeted functionalities. Traditional materials research and development primarily rely
on experimental trial-and-error and empirical knowledge. This process is not only time-consuming and
resource-intensive, but also often constrained by the limitations of human cognition, making it difficult to
meet the urgent demand for high-performance and functionally customized materials. In recent years, the
emergence of artificial intelligence (AI) for Material Science, particularly the breakthroughs in large language
models (LLMs) and AI fundamental models, has brought unprecedented opportunities to accelerate
materials discovery, reshaping the research paradigm of material science with unparalleled depth and
breadth .
[1-4]
The integration of AI with material science has followed a clear evolutionary path. The earliest attempts
involved expert systems, which encapsulated human knowledge into strict rules for specific tasks such as
phase diagram prediction or identification of organic molecules . While developing, their reliance on
[6]
[5]
manually organizing knowledge limited their scope and adaptability. The rise of machine learning (ML),
powered by increasing computational resources and materials data, marked a paradigm shift towards
data-driven discovery. ML models are good at discovering hidden patterns in high-dimensional spaces for
tasks such as property prediction and virtual screening. However, they are usually designed for specific tasks,
resulting in a lack of universality. Recently, breakthroughs in LLMs endowed them with transformative
capabilities in natural language understanding and generation, making it possible to interact with vast
scientific literature and knowledge bases and leverage the knowledge value they contain. On this basis, AI
agents represent a new research paradigm. Central to this transformation is the research paradigm shift from
AI for Material Science to Agentic Material Science . In this new paradigm, LLM-based agents are no longer
[7]
limited to specialized models for performing a single task; instead, they evolve into “virtual researchers”
capable of autonomous reasoning, planning, decision-making, and executing complex scientific research
workflows . AI agents are transforming material science from a data-driven to an agent-driven paradigm.
[8,9]

AI agents, with their powerful capabilities in natural language understanding and reasoning, multimodal
knowledge fusion, and professional integration with domain-specific tools, are becoming the critical
facilitators integrating the full workflow of materials research and development from concept and theory to
computation and experiment [10,11] . They are reshaping every stage of materials discovery. This agent-in-
the-loop paradigm ultimately aims to achieve fully autonomous self-driving laboratories (SDLs), enabling
end-to-end automation in literature mining, hypothesis generation, experimental design, materials synthesis,
and data interpretation. It signifies a rapid evolution of materials research and development toward complete
autonomy and intelligence .
[12]
Currently, the application of AI agents in material science is evolving along two distinct yet interconnected
pathways: materials design and materials synthesis and characterization . In materials design, agents have
[13]
demonstrated diverse applications including precisely extracting structured knowledge from large volumes of
scientific literature to facilitate knowledge discovery , enabling accurate and in-depth material knowledge
[14]
question answering to promote scientific communication, generating innovative material hypotheses to
guide material design, creating and optimizing material structures to explore vast material spaces, and
performing complex theoretical calculations to predict material properties [15,16] . These applications form a
comprehensive system for materials design, significantly enhancing research efficiency and innovative
capacity. In materials synthesis and characterization, the role of agents extends into experimental workflows
by enabling the exploration of novel synthesis processes beyond conventional limits, automating material