Page 120 - Read Online

Page 12 of 31 Li et al. J. Mater. Inf. 2026, 6, 10

multi-agent framework where an LLM-driven scientist agent generates molecules, while multiple analyzer
agents invoke RDKit tools for evaluation and feedback, achieving advanced levels in multiple molecular
optimization tasks . Organic Structure Directing Agent (OSDA) employs an LLM to generate candidate
[68]
structures, which are then evaluated by computational tools. Subsequently, the self-reflector summarizes and
provides feedback to guide the LLM in optimizing subsequent outputs . The MAPPS (Materials Agent
[69]
unifying Planning, Physics, and Scientists) framework further conducts workflow planning through LLM,
achieving enhanced stability and novelty in material design through integrating generation, physical force
field models, and scientist feedback .
[70]

In summary, AI agents for material structure design are rapidly evolving, demonstrating diverse technical
pathways. These systems have achieved notable results, generating hundreds of thousands of novel structures
ranging from MOFs and molecules to metamaterials, some of which have been successfully synthesized and
validated. The implementation of these systems mainly relies on the collaboration among different
components: specialized generative models (diffusion, GANs, RL, and LLMs) provide efficient exploration of
the vast material space; physics-based simulation tools ensure credibility and guide the search toward
reasonable structures; and LLMs act as the cognitive core, coordinating the entire workflow, facilitating
complex reasoning and multi-agent collaboration. This process follows the flow of
“generation-evaluation-optimization”, forming a closed loop and improving the design based on feedback.
The progress marks a paradigm shift from traditional targeted high-throughput screening toward intelligent
creation in materials design.

However, this field still faces numerous challenges at present. Firstly, the alignment between generation
intention and physical feasibility is insufficient, with some generated structures potentially violating
thermodynamic stability, synthetic feasibility, or chemical constraints of specific systems. Secondly, the
scarcity and bias of the training data limit the generalization and transferability of generation models among
different material systems. Existing systems fall short in modeling complex material systems with multi-scale
and multi-physical field coupling. Looking ahead, agent-driven material structure generation should place
greater emphasis on the deep integration of physical prior knowledge. By embedding physical knowledge
within the models, the generation of physical knowledge enhancement is achieved to improve the rationality
and innovation of the generated structure.

Property calculation
Materials property calculation is the cornerstone for validating the feasibility of materials design and deeply
understanding its performance. Although traditional computational methods, from ab initio calculations to
molecular dynamics and finite element analysis, have long been the foundation of theoretical insights, they
typically have high barriers to entry, require significant computational resources, and involve manual,
time-consuming workflows. AI agents demonstrate broad application potential in the field by addressing
these limitations. They can use material computational tools and trained ML models to achieve precise and
rapid calculations of material properties. This establishes a new paradigm that systematically enhances the
efficiency, accessibility, and scale of computational materials science.

A primary advancement of AI agents lies in democratizing complex computational tools. By natural
language interaction and modular collaboration, AI agents have been applied to material computational tool
empowerment and workflow automation. They can construct and automatically execute theoretical materials
property calculation workflows, significantly lowering the barrier to using professional material
computational tools. This allows researchers to rapidly invoke complex simulation processes via natural
language instructions, achieving efficient operation from setup to execution in materials property
calculations. These tasks traditionally require deep expertise in software-specific syntax and theoretical