Page 124 - Read Online

Page 16 of 31 Li et al. J. Mater. Inf. 2026, 6, 10

manufacturing, AutoMEX (a framework integrating LLMs as AI agents to automate the material extrusion
process) autonomously completes the entire workflow, including CAD model generation, slice code creation,
and 3D printing operations. This demonstrates its advantage in end-to-end automation of material
manufacturing processes . In the field of carbon nanotube synthesis, Carbon Copilot (CARCO) employed
[98]
chemical vapor deposition (CVD) experimental equipment to discover a Ti-Pt bimetallic catalyst within 43
days, achieving a synthesis precision of 56.25% for horizontally aligned carbon nanotube arrays . Materials
[99]
synthesis and characterization is a complex process full of uncertainties, involving intricate step planning,
process exploration, and data analysis. This section will explore how AI agents can be applied to various
aspects of process exploration, material characterization, and data analysis, analyzing their roles in key tasks
in material synthesis and characterization.

Synthesis process exploration
The first step in the automated material synthesis is to explore the material synthesis process. In this domain,
AI agents demonstrate multi-dimensional and adaptive optimization capabilities. By integrating prior
knowledge from knowledge graphs, autonomous exploration via RL, and multi-objective optimization
algorithms, they can efficiently balance multiple constraints such as performance, cost, and safety, to achieve
full autonomy in reaction route exploration, process parameter search, and multi-objective trade-offs. This
significantly enhances the efficiency of synthesis process design and parameter optimization for various
materials, effectively breaking through the bottlenecks of traditional process exploration that relies on
empirical trial-and-error, struggles with coupled parameters, and faces challenges in traversing complex
pathways.

Specifically, by integrating LLMs, knowledge graphs, and RL, agents achieve fully autonomous exploration
from reaction route recommendation and process parameter search to multi-objective trade-offs. For
example, the agent system based on LLM and knowledge graph automatically extracts reaction data from
literature, constructs retrosynthetic pathway trees, and combines multi-branch reaction path search
algorithms to recommend new synthesis processes for materials such as polyimides, overcoming the
limitations of traditional methods in designing reaction routes for complex macromolecules [100] .
Furthermore, for quantum material synthesis, an offline RL agent, trained through tens of thousands of
simulations, successfully predicts process parameters for synthesizing MoS single crystals via CVD,
2
demonstrating the predictive capability of agents in complex processes such as high-temperature and
multi-phase systems .
[101]
Material processes often require balancing multiple conflicting objectives such as performance, cost, and
safety, while satisfying engineering constraints. AI agents can address this challenge through advanced
algorithms to enable efficient decision-making. In the exploration space of process parameters, RL agent can
achieve the optimization of process parameters guided by target properties through interaction with the
environment [102,103] . To address the challenge of multi-objective cooperative optimization, a Bayesian
optimization algorithm agent efficiently explores the Pareto frontier among multiple objectives, such as
optical properties and reaction rates, in silver nanoparticle synthesis, significantly enhancing the feasibility of
synthesis processes under highly constrained conditions .
[104]
A key advancement lies in the integration of embodied intelligence, where AI agents bridge computational
decision-making and physical execution. By integrating real-time detection and perceptual feedback into the
experimental process, agents achieve closed-loop optimization. For example, the CLAIRify (LLM-based
natural language processing module) system employs constraint-aware motion planning to translate user
experimental descriptions into safely executable robot experimental actions. Through real-time sensor
detection and feedback, it successfully performs high degrees of freedom operations such as solubility