Page 118 - Read Online

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

The development of AI agents in materials hypothesis generation and design reflects functional
enhancement from expanding external capabilities to optimizing intrinsic reasoning mechanisms. This
progress indicates that agents are no longer merely engaged in simple data association, but begin to
systematically integrate multi-source knowledge, simulate human scientific inspiration and reasoning
processes, and thereby generate more reasonable and verifiable material hypotheses. By efficiently integrating
and associating originally fragmented knowledge units, agents not only improve the accuracy of hypothesis
generation but also significantly enhance the potential for successful application. However, it should also be
recognized that at present, the success rate, innovativeness and feasibility of generated hypotheses still
require extensive experimentation and algorithmic optimization for validation and consolidation. Therefore,
it is necessary to establish more refined evaluation frameworks for hypothesis quality and develop algorithms
capable of quantifying the success rate of hypotheses. Such advancements will not only enhance the creativity
of agents but also ensure that the scientific hypotheses have high verifiability and transformation value,
thereby driving materials design into an efficient and reliable intelligent era.

Structure design
Transforming qualitative material hypotheses into quantitative material structures is a critical leap in the
material design process. By integrating proprietary generative models and data-driven computational tools
into LLM-based workflows, AI agents enable automated and intelligent closed-loop exploration of material
structure generation. According to the technical approaches adopted for the generation module, existing
agent-driven material structure generation methods can be broadly categorized into two types.

The first category involves agents that employ specialized generative models for structure design, while LLMs
oversee regulation and control. The core of this approach lies in leveraging models such as diffusion models,
generative adversarial networks (GANs), or reinforcement learning (RL) to efficiently explore vast chemical
and structural spaces, with LLMs coordinating workflows, evaluating generated outputs, suggesting
optimization directions, and orchestrating multi-step processes. For instance, in MOF design, the MOFGen
employs an LLM agent to propose new compositions, a diffusion model to generate crystal structures, and
combines quantum mechanics tools with a synthesis feasibility assessment agent for screening and
optimization. This approach has generated hundreds of thousands of novel MOF structures and successfully
synthesized five MOF materials, validating the effectiveness of the agent . CrossMatAgent integrates
[57]
generative models such as DALL-E 3 and Stable Diffusion XL to create metamaterial structural patterns, with
a GPT-4o-driven multimodal agent responsible for performance analysis and supervisory feedback, thereby
automating the generation of metamaterial structures . LIDDIA (Language-based Intelligent Drug
[58]
Discovery Agent) employs molecular generation models guided by LLM reasoning to steer the molecular
generation process. By balancing chemical space exploration with optimal molecular utilization, it
successfully generates molecules that meet key metrics . Specialized generative models based on RL have
[59]
been applied in the ACEGEN framework, a comprehensive and streamlined toolkit for generative drug
design , and in CSstep, a drug molecule generation and optimization framework based on Markov decision
[60]
processes and reinforcement learning . By applying modular design and multi-agent decision-making
[61]
strategies, the molecular generation process can be progressively optimized and rendered more interpretable.
In the design of two-dimensional (2D) material structures, an RL agent can efficiently generate the MoS₂
structures with enhanced tensile strength after being trained with a small amount of data . ChatMOF
[62]
utilizes LLM to process natural language inputs and combines databases and ML tools to achieve MOF
prediction and generation . MatAgent is a generative framework for inorganic materials discovery that
[63]
leverages the reasoning capabilities of large language models. In inorganic materials design, it improves the
effectiveness, uniqueness, and novelty of candidate materials through iterative feedback between generation
and property prediction models [Figure 4A]. The ProtAgents platform deploys multiple function-specific
[64]
agents responsible for knowledge retrieval, structural analysis, physical simulation, and result analysis to
design proteins with target performance .
[65]