Page 125 - Read Online

Li et al. J. Mater. Inf. 2026, 6, 10 Page 17 of 31

measurement and recrystallization in material synthesis . Similarly, K-Agents uses LLM-based planning to
[105]
model multi-step experimental operations as robot actions, and calibrates them in real-time using knowledge
of laboratory operations and sensor detection feedback . The integration of digital twin technology further
[106]
enhances this paradigm by providing virtual models of physical systems. These simulate robot trajectories,
predict collisions, and optimize task sequences before physical execution. Digital twins also support real-time
monitoring and adaptive reconfiguration, improving efficiency and resilience. For instance, as a fundamental
support platform, LabUtopia provides a high-fidelity simulation environment for deploying AI agents in
physics laboratory settings. It integrates the Simulation Lab (LabSim) simulator, which models
high-precision physicochemical interactions, LabScene for generating diverse scientific scenarios, and the
LabBench benchmark, covering tasks ranging from atomic actions to long-term operations. By offering a
simulation environment that supports multiple physicochemical interactions, including over 200 scenarios
and instrumental assets, LabUtopia enables large-scale training and evaluation for 30 categories of complex
experimental tasks. It thus serves as a standardized platform for training and comprehensively assessing the
perception, planning, and control capabilities of scientific embodied agents in complex laboratory settings,
laying a solid foundation for developing embodied AI agents capable of long-term planning and
sophisticated operations in real laboratories [107] . In dangerous or fully autonomous environments, safety
protocols, including safety barriers, collision detection sensors, and fail-safe mechanisms, are indispensable.
By combining motion planning, digital twin simulations, and robust safety systems, robot platforms enhance
laboratory efficiency while ensuring safe and effective operations .
[108]
At the system level, by introducing multi-robot scheduling, hierarchical control, and modular planning,
agents can achieve system-level resource optimization in experimental design, significantly improving
experimental efficiency. A multi-robot and multi-task scheduling system employing constraint programming
methods optimizes task allocation among 3 robots and 18 experimental stations while supporting dynamic
task insertion. This reduces the total execution time of complex experiments by nearly 40%, enhancing the
parallelism of high-throughput experimentation .
[109]

These advances indicate that AI agents are transforming synthesis process exploration from trial-and-error
into a data-driven science. They establish a new paradigm for process exploration that encompasses path
design, parameter search, and multi-objective optimization, providing the core support for the
transformation of materials synthesis processes from experience to precise design. In addition, this direction
also faces several challenges. Firstly, due to the difficulty in accurately simulating the physicochemical effects
under highly complex or extreme conditions, the gap between simulations and real physicochemical
processes is still significant. This may influence the generalization ability of the agent trained in a simulation
environment when applied to real-world laboratory settings. Furthermore, the stability, safety, and
interpretability of system during long-term autonomous operation in laboratories require further
enhancement to build more reliable, transparent, and continuously upgradable material agents. These
challenges highlight the direction for research on the next generation of AI agents for material synthesis.

Materials characterization
The properties of synthesized materials need to be confirmed through characterization. Experimental
characterization is the core means to determine whether the synthesized product meets expectations.
Through techniques such as X-ray diffraction (XRD), scanning electron microscopy (SEM), and
transmission electron microscopy (TEM), information on the material’s crystal structure, microstructure,
and composition distribution is obtained, thereby validating the rationality of process parameters. By
integrating ML models, experimental equipment control logic, and multimodal data parsing capabilities, AI
agents have significantly addressed bottlenecks in traditional characterization, such as labor-intensive
manual operations, time-consuming data analysis, and difficulties in identifying complex systems. Their