Page 123 - Read Online

Li et al. J. Mater. Inf. 2026, 6, 10 Page 15 of 31

computational simulations serve as a rigorous, physics-based validator, a function absent in traditional
workflows. By incorporating computational verification and feedback, the reliability and scientific rationality
of LLMs are enhanced. For example, LLaMP employs RAG and a hierarchical ReAct agent to dynamically
query databases such as Materials Project and first-principles calculation tools, correcting LLM prediction
deviations for properties such as bulk modulus and bandgap, thereby significantly reducing hallucinations in
material property predictions . PolySea, a specialized large model for polymers, needs to undergo triple
[91]
verification through the polymer genome database, GNN model, and DFT calculation after generating new
structures to ensure the credibility of the design results . AtomAgents integrates multimodal data and
[92]
physical simulations through dynamic multi-agent collaboration, embedding physical principles of alloys
into the design process, which improves both the accuracy of performance predictions and the
interpretability of alloy design . ChemHAS (Chemical Hierarchical Agent Stacking) enhances the
[93]
robustness of chemical property prediction by using the agent-stacked structure to compensate for the
prediction errors of chemical tools with data . HoneyComb combines a high-quality materials science
[94]
knowledge base with an adaptive tool selection module, significantly improving accuracy in domain tasks by
matching and using the most relevant knowledge and computational tools . xChemAgents employs a
[95]
dual-agent collaboration mechanism with Selector and Validator. The former adaptively screens chemical
descriptors relevant to target properties and provides natural language explanations, while the latter
iteratively introduces physical constraints into the descriptors, reducing the mean absolute error of
multimodal property predictions by 22% . This integration enables AI-generated hypotheses to be validated
[96]
against physical laws, creating a self-correcting and trustworthy design loop, laying an important foundation
for establishing scientifically rigorous, domain-specific LLMs.

The application of AI agents in materials property calculation plays a dual role of empowerment and
feedback. It simplifies and automates complex computational simulations, thereby enhancing research
efficiency; meanwhile, the results from these simulations serve as verification feedback, improving the
scientific rigor and credibility of the agents’ outputs. This establishes a solid and reliable theoretical
foundation for the eventual verification, accelerating the entire journey from conceptual design to
experimental realization. There are also some challenges in the widespread adoption and further
development of AI agents for materials property calculation. A major issue is the accuracy and generalization
of data-driven ML models, which highly depend on the quality, quantity, and diversity of training data.
Furthermore, the multiscale characteristic of materials brings significant modeling challenges, as seamlessly
integrating simulations from the quantum scale to the macroscopic scale within an agent framework remains
an issue. The high computational cost of high-precision simulations, such as DFT, also limits the scale and
speed of agent-driven closed-loop exploration.

In summary, AI agents have demonstrated powerful comprehensive capabilities in the materials design stage
from literature data mining, knowledge question answering, and hypothesis generation, to specific materials
structure design and property calculation. They have achieved a closed-loop, autonomous materials design
from knowledge input to theoretical validation.

AGENTIC MATERIALS SYNTHESIS AND CHARACTERIZATION
Materials synthesis and characterization are the key steps in transforming theoretical designs into physical
substances. By integrating robot control and experimental equipment, AI agents are being deeply integrated
into material synthesis and characterization, enabling the automation and intellectualization of autonomous
material synthesis and characterization workflows, thereby significantly enhancing experimental efficiency
and exploratory capabilities. For instance, in the development of high-entropy alloy catalysts, agents utilizing
Joule heating technology achieve high-throughput synthesis and performance screening. This demonstrates
the powerful ability of agents to accelerate material discovery using experimental equipment . In additive
[97]