Page 96 - Read Online

Wang et al. J. Mater. Inf. 2026, 6, 9 Page 3 of 15

Agent performs rational inverse molecular design by extracting domain knowledge from scientific literature
and iteratively refining candidate compounds using surrogate models . Similarly, symbolic regression
[34]
frameworks have shown that LLM-based Agents can uncover interpretable material laws, such as descriptors
for the glass-forming ability of metallic glasses . For materials data extraction and synthesis planning, the
[35]
Eunomia Agent uses zero-shot learning to convert unstructured scientific text into structured datasets, while
the MOFsyn Agent accelerates experimental workflows by optimizing metal-organic framework (MOF)
synthesis protocols through natural language interaction and mechanism-aware reasoning . However, these
[36]
agents are typically designed for defined application scenarios, with relatively linear workflows. In contrast,
materials design often requires significantly more complex workflow planning, involving the integration of
quantum chemistry, molecular dynamics, and high-throughput simulations under diverse constraints and
objectives - posing greater challenges to the reasoning and orchestration capabilities of LLM-based systems.
Moreover, the highly repetitive nature of constructing similar workflows across different systems
underscores the need for agents that can generalize workflow design patterns and alleviate researchers from
the burden of manually configuring routine processes. Realizing such capability would not only enhance
automation and scalability but also allow domain experts to focus more on scientific insight and
decision-making.

To address this, we present an LLM-based collaborative materials design Agent, empowered by sequential
reasoning and the Reasoning and Action (ReAct) framework . By seamlessly integrating LLMs with a suite
[39]
of computational materials science tools and databases, the Agent constructs simulation workflows,
dynamically refines task plans in response to intermediate results, and establishes a fully closed-loop research
and development system. The Agent supports a broad spectrum of computational tools, including
first-principles simulation engines, molecular design and screening platforms, and ML models, thereby
enabling unified handling of both periodic crystalline systems and molecular compounds. Using Model
Context Protocol (MCP) , the Agent can flexibly invoke the most appropriate tools based on task
[40]
objectives, accuracy requirements, and computational constraints. Its modular and extensible architecture
allows seamless adaptation to a wide range of materials design tasks, including band structure prediction,
molecular property screening, structure generation, and multi-objective optimization. We validate the
generality and effectiveness of the Agent through representative case studies in both inorganic and organic
domains, such as the electronic structure evaluation of two-dimensional (2D) semiconductors and the
inverse design of electrolyte additives targeting specific energy levels. These results demonstrate the Agent’s
ability to make context-aware decisions, adapt its strategy through real-time feedback, and efficiently
converge toward optimal solutions across diverse materials spaces. Taken together, this work proposes a
scalable paradigm for intelligent materials design - reducing human intervention while enhancing
adaptability, reusability, and design efficiency - and demonstrates its application in the design of 2D
semiconductors and electrolyte additives.

MATERIALS AND METHODS
To support intelligent and adaptive materials design, we developed an AI-Agent system, Material Buddy
(Matty), capable of autonomously constructing and executing complex simulation workflows. As illustrated
in Figure 1, Matty is built upon LLM-driven architecture that tightly couples sequential reasoning with
ReAct. At the core of the system is the Qwen3-235B-A22B-FP8 LLM, which serves as the cognitive engine
[41]
for natural language understanding, task decomposition, and strategic planning. Matty also adopts the MCP,
a standardized interface that facilitates seamless integration of diverse computational tools. Leveraging the
MCP service, Matty can dynamically plan, invoke, and execute tasks in a stepwise fashion by autonomously
translating high-level material design objectives into actionable workflows. Built on this LLM + MCP