Page 129 - Read Online

Li et al. J. Mater. Inf. 2026, 6, 10 Page 21 of 31

1.37 million candidates, it identified eight top-20 high-performance materials using only 200 evaluations,
highlighting the advantage of agents in analyzing structure-property relationships from experimental data .
[128]
It can be seen from these that in the experimental optimization, agents not only achieved a leapfrog
improvement in performance, but also revealed the key factors and regulatory mechanisms that affect
material performance in a data-driven way.

In summary, AI agents transform data analysis from an isolated, post-hoc step into a real-time feedback loop
in the experimental process. They efficiently parse high-dimensional materials data, dynamically integrate
knowledge, and perform cross-scale reasoning. This enables the discovery of hidden patterns and novel
mechanisms from complex datasets, the proposal of verifiable new hypotheses, and the optimization of
experimental pathways. Consequently, they provide direct, data-driven decision support for materials science
research, significantly improving experimental efficiency and scientific discovery.

This section demonstrates the deep integration of AI agents throughout the entire process of materials
synthesis and characterization. From the exploration of material synthesis processes, automated
characterization to data analysis and conclusion summary, agents are shifting the paradigm of materials
synthesis from “experience-driven” to “intelligence-driven”. They not only significantly enhance
experimental efficiency but also make it possible to “synthesize on demand” materials with targeted
properties.

AGENTIC MATERIALS CREATION
As the capabilities of AI agents in materials design and synthesis continue to mature, the trend of their
development is to form SDLs. The SDL is the ultimate embodiment of materials AI agents, integrating the
entire workflow for materials creation from design and synthesis to validation. Through a trinity architecture
of “hardware-software-agent”, it deeply integrates LLMs, MAS, and robotics technologies. This enables a
fully autonomous closed loop from experimental design, execution, to analysis and optimization, thereby
transforming the paradigm of materials creation in unprecedented ways.

The development of agent-driven SDLs is a progressive evolution, clearly reflecting the synergistic
advancement of AI and robotics technologies. Early groundbreaking work, such as the mobile robotic
chemist, broke the constraints of traditional fixed workstations. By operating freely in a standard laboratory
and utilizing batched Bayesian search algorithms, it autonomously conducted 688 photocatalyst experiments
in eight days and screened out a hydrogen evolution photocatalyst with sixfold higher activity,
demonstrating the feasibility of deploying automated researchers rather than mere automated
instruments . A-Lab, an autonomous laboratory for the solid-state synthesis of inorganic powders, further
[129]
achieves autonomous synthesis of inorganic powders by leveraging computational predictions, literature
data, and ML. Within 17 days, it successfully synthesized 41 novel compounds, highlighting the potential of
the integration of AI and robotics to accelerate materials discovery [130] [Figure 7A]. Subsequently, the flow
chemistry automation platforms combined liquid processors, syringe pumps, adjustable continuous-flow
photoreactors, and in-situ NMR spectroscopy, and achieved automated optimization and scale-up of
photocatalytic reactions through closed-loop Bayesian optimization strategies .
[131]

With the development of LLMs, the deep empowerment of LLMs has driven the evolution of AI agents from
performing single functions to handling multifaceted tasks, significantly enhancing the cognitive and
decision-making capabilities of SDLs. For instance, systems such as Coscientist (a multi-LLMs-based
intelligent agent) [132] and ChemCrow (an LLM-based chemistry agent for tasks in organic synthesis, drug
discovery, and materials design) [133] used LLMs as core planners. These agents autonomously perform