Volume 6, Issue 1 (2026) – 17 articles
Cover Picture: This cover illustrates the central theme of this Review: the agent-driven paradigm in material science, in which AI agents autonomously orchestrate the entire workflow of materials discovery. At the center is an AI agent that serves as the cognitive core, coordinating every stage of materials research. The three surrounding bubbles represent the agent's capabilities: the neural network at the top represents the various models invoked by the agent for tasks such as property prediction and structure generation; the large language model in the middle supports the agent with natural language reasoning, literature mining, and tool invocation; and the periodic table at the bottom symbolizes the vast materials database that the agent can retrieve and integrate. The three scenarios extending outward from the agent collectively form a closed-loop, agent-driven materials research process. The materials structure on the right illustrates the agent's application in material design and property prediction. The characterization spectrum on the left represents the autonomous collection and intelligent analysis of instrumental data. The robotic arm in the bottom right symbolizes the self-driven laboratory, where hypothesis generation, synthesis, and validation are executed end-to-end without human intervention. Together, these elements visualize how AI agents seamlessly integrate knowledge, computation, and experimentation, thereby accelerating the advancement of materials science towards a truly agent-driven paradigm.
view this paper 
Back Cover Picture: Designing new materials traditionally requires extensive expertise in computational methods, specialized software, and workflow engineering, creating a high barrier for researchers. Here, we introduce Matty (Material Buddy), an AI-driven autonomous agent that enables end-to-end materials design and simulation through natural language interaction. Built upon a large language model, Matty integrates perception, memory, planning, execution, and learning modules into a unified framework connected via a standardized tool interface.
Unlike conventional approaches, Matty can automatically interpret user instructions, decompose complex tasks, and orchestrate diverse computational tools, including density functional theory, molecular dynamics, and machine learning models. In benchmark demonstrations, Matty successfully performs automated electronic structure calculations for two-dimensional materials and conducts inverse molecular design under multiple property constraints, achieving results consistent with literature while requiring no manual scripting or parameter tuning.
Furthermore, Matty incorporates a closed-loop learning mechanism, continuously improving its performance by integrating newly generated data. This work highlights a paradigm shift from tool-assisted computation to autonomous AI agents, significantly lowering the barrier to materials discovery and paving the way for intelligent, scalable, and user-friendly materials design systems.
view this paperUnlike conventional approaches, Matty can automatically interpret user instructions, decompose complex tasks, and orchestrate diverse computational tools, including density functional theory, molecular dynamics, and machine learning models. In benchmark demonstrations, Matty successfully performs automated electronic structure calculations for two-dimensional materials and conducts inverse molecular design under multiple property constraints, achieving results consistent with literature while requiring no manual scripting or parameter tuning.
Furthermore, Matty incorporates a closed-loop learning mechanism, continuously improving its performance by integrating newly generated data. This work highlights a paradigm shift from tool-assisted computation to autonomous AI agents, significantly lowering the barrier to materials discovery and paving the way for intelligent, scalable, and user-friendly materials design systems.
