Page 126 - Read Online

Page 18 of 31 Li et al. J. Mater. Inf. 2026, 6, 10

applications have covered multiple characterization technologies, including optical microscopy, scanning
tunneling microscopy (STM), atomic force microscopy (AFM), XRD, SEM, and TEM, and have achieved
efficient characterization in diverse systems such as 2D materials, zeolites, alloys, and organic crystals. By
rapidly identifying phases and analyzing microstructures, they provide key support for materials discovery.

For instance, in image recognition and analysis, the 2D-TLK model is a lightweight framework for layer
identification in 2D materials, integrating a FastViT encoder with a composite decoder combining Lite
Reduced Atrous Spatial Pyramid Pooling and Kernels net. Combined with a MAS, it enables high-precision
identification and natural language interpretation of thickness and size in 2D materials such as MoS , WS ,
2
2
and graphene using optical microscopy. The model achieves an accuracy of 95.48% and demonstrates strong
generalization capability and real-time reasoning performance, effectively solving the problem of traditional
optical contrast methods that are susceptible to substrate and illumination interference [110] . In the field of
STM, a deep learning-based agent system can autonomously perform image quality assessment, surface
recognition, and probe adjustment in STM. It enables large-area continuous measurement and molecular
statistics under low-temperature conditions (78 K), breaking through the limitations of conventional STM
that rely on manual probe adjustment and struggle with long-term stable characterization, thereby
significantly enhancing data throughput and reliability [111] . Similarly, in AFM applications, the LLM-driven
Artificially Intelligent Lab Assistant (AILA) MAS can coordinate tasks such as calibration, feature detection,
mechanical property measurement, and layer identification . These advancements collectively demonstrate
[112]
how agents transform processes that are heavily dependent on humans, subjective, and low-throughput into
automated, intelligent, and repeatable workflows. By overcoming traditional limitations such as operator bias
and environmental interference, agents not only enhance measurement accuracy but also enable large-scale,
statistically meaningful microstructural analysis.

In the field of crystal structure analysis, AI agents offer new pathways for the automated interpretation of
XRD data. For instance, the Crystallography Companion Agent integrates a CNN model trained on
physically accurate synthetic datasets, overcoming the overconfidence issue of traditional neural networks by
outputting probabilistic phase classifications. This significantly enhances the accuracy and efficiency of phase
identification in both organic and inorganic materials, reducing the analysis time from several weeks to just a
few hours . CrystalShift combines symmetry-constrained optimization and best-first tree search to enable
[113]
probabilistic phase labeling and structural parameter extraction from high-throughput XRD data. Without
additional phase space information, it can output phase combination probabilities, outperforming existing
methods and demonstrating strong robustness in experimental data. Moreover, it can be seamlessly
integrated into synchrotron radiation high-throughput workflows [Figure 6]. In addition, in the fields of
[114]
electron diffraction and scanning transmission electron microscopy (STEM) characterization, SerialRED
technology automates data acquisition and hierarchical clustering analysis, enabling high-throughput phase
analysis of complex polycrystalline mixtures with only a small amount of samples. It can even detect
ultra-low content phases that are undetectable by XRD [115] . Meanwhile, an ensemble learning iterative
training (ELIT) workflow focuses on ML automation for STEM characterization. By developing a universal
language to adapt to various STEM devices, it dynamically adapts to changing imaging conditions, increasing
the detection accuracy of atomic positions to over 90%. It also supports in-loop validation with theoretical
simulations for structure-property correlation analysis [116] . These agents address one of the most persistent
challenges in materials science: reliably and rapidly decoding complex multiphase systems without relying
heavily on expert intuition or pre-existing phase databases. By reasoning and handling experimental noise
robustly, these systems elevate the objectivity and scalability of structural analysis to a new level. This not
only shortens analysis time, but also enables researchers to explore more complex components and discover
previously overlooked materials.