Page 116 - Read Online

Page 8 of 31 Li et al. J. Mater. Inf. 2026, 6, 10

Hypothesis generation
The core of materials design lies in proposing novel and reasonable hypotheses. The reasoning capabilities
demonstrated by AI agents in knowledge question answering can evolve into creative thinking, enabling
them to transform fragmented knowledge into verifiable material hypotheses and shifting materials design
from “trial-and-error” to “targeted creation” [38,39] . Research strategies for hypothesis generation by agents can
be broadly divided into two categories. One focuses on augmenting the agent’s abilities by integrating
external tools and knowledge sources, while the other emphasizes optimizing the internal collaboration
mechanisms and reasoning strategies to enhance the quality and innovativeness of generated hypotheses.
These approaches jointly advance the automation and rationalization of the materials design process.

In terms of knowledge integration and tool utilization, agents make up for inherent limitations by using
external tools and structured knowledge, thereby overcoming the constraints of a single model. For instance,
dZiner combines first-principles calculations with physics-inspired models to iteratively optimize material
designs, achieving the rational design of MOFs from target properties to structures [Figure 3A]. SciAgents
[40]
integrates large-scale knowledge graphs with data retrieval tools to explore interdisciplinary connections in
the field of biomimetic materials, autonomously generating and optimizing material design hypotheses .
[41]
ResearchAgent employs a reviewer agent to iteratively define new problems, propose methodologies, and
design experiments . SciAgent enhances the LLM with tools specifically built for scientific reasoning tasks,
[42]
enabling it to retrieve, understand, and apply tools to solve scientific problems . Dolphin establishes a
[43]
closed-loop research framework in which agents generate new ideas based on experimental feedback and
relevant literature, then autonomously implement experiments, debug code, and analyze results, enabling
continuous optimization of the research process . The plant-inspired materials design framework integrates
[44]
BioinspiredLLM, a large language model for the mechanics of biological and bio-inspired materials, together
with RAG tools and hierarchical sampling strategies. This approach enables effective extraction of
structure-property relationships from plant science literature and guides the experimental preparation of
novel pollen-based adhesives . These approaches essentially augment agents with toolchains and knowledge
[45]
bases, equipping them to function as super-experts.

In addition to empowering agents through the integration of external resources, another research focuses on
enhancing the intrinsic reasoning and discovery capabilities of agent systems by innovating in multi-agent
collaboration frameworks or reasoning-iteration optimization strategies, thereby further improving the
quality of generated hypotheses. This type of work emphasizes either interaction among multiple agents or
the self-improvement of individual agents to achieve more rigorous and creative scientific reasoning. For
example, ReConcile adopts a round-table discussion framework, where multi-round debates and
confidence-weighted voting mechanism help reach a consensus, significantly improving reasoning quality .
[46]
MOOSE-Chem employs a mathematical decomposition framework to break down hypothesis generation
into three subtasks: inspiration retrieval, hypothesis combination, and ranking, successfully reproducing
high-level chemical and materials hypotheses recently proposed by human scientists [Figure 3B].
[47]
MOOSE-Chem2 further formalizes fine-grained hypothesis discovery as a combinatorial optimization
problem. Through a hierarchical search method, it gradually refines details from general concepts to specific
experimental configurations, and discovers the optimal detailed hypothesis that can be immediately put into
experiment by smoothing the internal reward landscape . To more reasonably evaluate hypotheses,
[48]
MOOSE-Chem3 develops an experimental simulation feedback tool. By introducing a hypothesis ranking
task guided by simulated outcomes, candidate schemes were prioritized through clustering and simulation
results . AgentRxiv establishes a shared platform for agents that simulates the real-world scientific
[49]
community, enabling agents to iteratively advance their research based on each other’s research
achievements . PiFlow models scientific discovery from an information-theoretic perspective as a
[50]
structured uncertainty reduction process under the constraints of scientific principles, achieving hypothesis