Transforming TEM Adaptiveness for Transformative Materials Development

Abstract

Transmission electron microscopy (TEM) has long been driven by the pursuit of higher spatial resolution, culminating in sub-angstrom imaging capabilities that have revolutionized atomic-scale materials characterization. However, a fundamental disconnect has emerged between the idealized conditions required to achieve this ultimate resolution and the complex, dynamic, and reactive nature of real-world materials central to transformative technologies. This perspective proposes a paradigm shift for TEM development: transitioning from a singular focus on a “Resolution Revolution” to a broader “Adaptiveness Revolution.” This new paradigm encompasses two interconnected dimensions. Material Adaptiveness aims to extend TEM's characterization capabilities to diverse materials and operational environments through three key advances: (i) full-information imaging via multi-signal fusion and advanced algorithms; (ii) spatiotemporal multi-dimensional imaging to capture dynamic processes from seconds to femtoseconds across three-dimensional space; and (iii) in-situ/operando imaging under complex, multi-field coupled conditions. Knowledge Adaptiveness seeks to democratize access to advanced TEM through (i) encapsulation of expert knowledge into automated workflows; (ii) automation and hierarchical screening for high-throughput analysis; and (iii) remote operation with intelligent scheduling to optimize utilization of scarce instrumentation. We illustrate the critical need for this revolution by showcasing advanced TEM applications in three transformative material systems, including functional ceramics, high-strength structural materials, and energy catalytic materials. The synergistic development of Material and Knowledge Adaptiveness will not only deepen fundamental understanding of material behavior under realistic conditions but also accelerate the discovery and deployment of transformative materials to address global challenges in energy, environment, health, and information.

Graphical abstract

Keywords

Transmission Electron Microscopy / Paradigm Shift / Material Adaptiveness / Knowledge Adaptiveness / Transformative Materials

Highlight

	● Proposes shifting TEM focus from higher resolution to an "Adaptiveness Revolution" for real word transformative materials.
	● Material Adaptiveness expands TEM capabilities for diverse materials, multidimensions and complex environments.
	● Knowledge Adaptiveness transfers expert knowledge into AI and automation, democratizing access for the wider research community.

1 Introduction

Since the advent of the transmission electron microscope, a primary thread of its development has been the relentless pursuit of higher resolution. From the early visualization of crystal defects to the successful commercialization of aberration correctors that pushed real-space resolution into the sub-angstrom regime, hardware innovations have played an indispensable role [1]. In recent years, the rise of computational microscopy methods, exemplified by electron ptychography, has enabled the extraction of fine phase information from experimental data through algorithms, approaching the fundamental imaging limits imposed by atomic thermal vibrations [2,3]. This revolution has allowed humanity to intuitively “see” atoms, laying the cornerstone for atomic-scale research in fields such as materials science, condensed matter physics, chemistry, and structural biology.

However, alongside this progress in resolution, a contradiction has emerged: the ultimate resolution achieved under ideal laboratory conditions on model samples (e.g., ultra-thin, stable, clean, radiation-resistant crystals) is often difficult to replicate in the study of numerous real, complex material systems [4]. The solid-liquid interfaces in high-performance batteries, the superalloys in turbine blades, the nano-heterostructures in semiconductor devices, and the organic-inorganic hybrid perovskites in photovoltaics—these material systems, central to transformative technologies, exhibit complex intrinsic characteristics precisely outside the ideal imaging conditions of conventional TEM: electron beam sensitivity, environmental dependence, dynamic evolution, and multi-field coupling. The mismatch between theoretical ultra-high resolution and practical application signals that the marginal benefit of purely pursuing higher resolution has significantly diminished.

Concurrently, the scientific quest to understand the origins of material functionality has deepened. The evolutionary behavior of multi-dimensional order parameters—lattice, charge, orbital, and spin—and their coupling under multi-field stimuli (mechanical, thermal, gaseous, electrical, optical) is increasingly recognized as the key to understanding the properties of transformative materials [5]. This demands that TEM technology not only “see” atomic positions but also elucidate the intrinsic link between dynamic structural evolution and functional response under the material's actual operating conditions.

Against this backdrop, this article proposes that TEM development needs a paradigm shift oriented towards transformative materials research (Fig. 1): from a “Resolution Revolution” to an “Adaptiveness Revolution.” This new paradigm comprises two interconnected core dimensions: Material Adaptiveness, which aims to overcome current TEM limitations concerning sample properties, state, and environment, expanding its capability to characterize fine structures and order parameters across a broader range of material systems and reveal atomic-scale information under near-realistic service conditions; and Knowledge Adaptiveness, which seeks to encapsulate and transfer expert knowledge through intelligent, automated, and remote operation, enabling non-specialist researchers to conveniently utilize advanced TEM techniques while optimizing the global allocation and output efficiency of scarce instrumentation (Fig. 2).

This article will first elaborate on the conceptual framework of the TEM Adaptiveness Revolution. Subsequently, using three classes of transformative materials—advanced functional ceramics, high-strength aerospace materials, and multi-purpose energy catalytic materials—as examples, it will discuss the current state of advanced TEM applications in materials science and outlook how the TEM Adaptiveness Revolution will empower key breakthroughs in these fields.

2 The TEM Adaptiveness Revolution

2.1 Material adaptiveness: expanding the boundaries of characterization

The core objective of Material Adaptiveness is to enable TEM to flexibly, robustly, and efficiently adapt to the diverse intrinsic characteristics of materials and complex operating environments. This requires breakthroughs on the following three levels:

2.1.1 Full-information imaging

The electron beam in a TEM, upon interacting with a sample, generates not only elastically scattered electrons at various angles but also stimulates secondary electrons (SE), characteristic X-rays, cathodoluminescence, Auger electron spectroscopy (AES), and other signals carrying material-specific information[6]. Traditionally, these signals are collected and analyzed in isolation, failing to achieve full-information utilization. Concurrently, the electron beam can induce adverse effects such as atomic displacement, bond breakage, sample heating, and charge accumulation, posing major obstacles to applying TEM to beam-sensitive materials like soft matter, organic materials, and metal halides [7]. To advance Material Adaptiveness, future efforts must enhance the comprehensive utilization of signals generated by electron-matter interactions while effectively mitigating electron irradiation damage.

At the hardware level, direct electron detectors (DEDs) are progressively replacing conventional charge coupled device (CCD) and complimentary metal-oxide-semiconductor (CMOS) cameras. By compressing the electron-photon-electron signal conversion process into direct electron readout, these detectors achieve a detective quantum efficiency of more than 90% and imaging frame rates of 120 kHz, making atomic-resolution imaging of beam-sensitive materials more feasible under low-dose conditions [8,9]. Their successful application in structural biology has significantly improved the resolution of protein macromolecules and has driven atomic-scale imaging of real-space structures in sensitive materials like metal-organic frameworks [10,11]. However, DEDs still face a trade-off between detection precision, imaging speed, and operating voltage; achieving an optimal balance among these three simultaneously is a future challenge. Furthermore, exploring non-traditional detector geometries and optical configurations can enable efficient collection of signals from different spatial regions, boosting detection efficiency. For instance, in energy dispersive X-ray (EDX) detectors, optimizing the detection area and geometry of silicon drift detectors (SDDs) and their placement within the pole-piece gap has allowed dramatically increasing of the solid collection angle from <1 steradian (sr) to more than 4 sr [12]. This improvement directly enhances imaging throughput by 4 times, making atomic-resolution elemental mapping, previously hindered by sample drift, feasible, such as studying elemental distribution in high-entropy alloys [13]. Brown et al. designed a square condenser aperture plate to match a square detector, departing from the traditional circular beam shape, thereby increasing electron beam utilization by 70% [14]. Song et al. proposed a hollow-cone pixelated direct electron detector capable of simultaneously performing high-precision electron ptychography reconstruction and electron energy loss spectroscopic (EELS) signal acquisition, paving the way for coupling lattice structure imaging with chemical, valence, and even phonon mapping [15].

At the algorithmic level, developing multi-signal fusion imaging is crucial to overcome the physical detection limits of single modalities and synergistically enhance multi-modal image quality through advanced image processing techniques. For example, Schwartz et al., by fusing annular dark field (ADF), EDX, and EELS signals, reduced the required dose to 1% of that needed for conventional high-resolution tomography, achieving 1 nm three dimensional (3D) spatial resolution and differentiation of elemental species in materials like Au-Fe₃O₄ metamaterials within an organic ligand matrix, Co₃O₄-Mn₃O₄ core-shell nanocrystals, and ZnS-Cu₀.₆₄S₀.₃₆[16]. While current multi-modal fusion is often limited to qualitative elemental analysis, the gradual integration of ptychography’s high spatial resolution, EDX’s quantitative elemental analysis, EELS’s valence state mapping, differential phase contrast’s (DPC) charge distribution mapping, momentum-resolved EELS’s phonon imaging, SE imaging for surface-sensitive atomic-scale topography and electronic structure, and Auger electron spectroscopy (AES) for nanoscale chemical state analysis holds immense promise for advancing research into the atomic and electronic structural origins, especially in deciphering the mystery of the origin of superconductivity [17–21]. Additionally, integrating image processing algorithms from computer vision and deep learning will significantly enhance information extraction and quantitative analysis from TEM images. Although noise sources in TEM are complex and diverse—including quantum noise from electron counting statistics, colored noise from inelastic scattering and chromatic aberration, and specific local artifacts from beam instability, detector response, or environmental interference—unsupervised or self-supervised deep learning denoising algorithms have demonstrated excellent performance [22–25]. Further combining these with few-shot learning frameworks like graph neural networks, or introducing physics-informed generative artificial intelligence (AI), can enable automated identification and statistical analysis of physical features in TEM images, such as vacancy defects in 2D materials or supported species in heterogeneous catalysts at the nanoscale [26,27]. This will greatly enhance the efficiency and precision of such analyses, laying the groundwork for high-throughput atomic characterization and rapid iterative research and development (R&D). By combining hardware and algorithmic breakthroughs, advanced imaging on difficult samples, including beam sensitive materials, will become more accessible, and Knowledge Adaptiveness (Section 2.2) will further democratize these techniques.

2.1.2 Spatiotemporal multi-dimensional imaging

The properties of many materials depend on atomic-scale spatiotemporal dynamic processes, which cannot be fully captured by static 2D projections alone. For porous materials like zeolites, the distribution of guest species within sub-nanometer channel networks directly dictates catalytic performance, but the complex channel structure makes intuitive interpretation from 2D projections difficult [28]. To capture such complexities, it is necessary to break the ambiguities of 2D projections in spatial interpretation by developing 3D imaging techniques. Currently, Miao’s group have achieved quantitative 3D coordinate analysis of every atom in crystalline nanoparticles like Au and FePt, and have extended this to amorphous nanoparticles like Si nanoparticles [29,30]. Future developments in characterization techniques and reconstruction algorithms will extend 3D atomic-resolution imaging to radiation-sensitive materials and significantly broaden the observable material volume. Furthermore, by integrating with full-information imaging, this will enable the fusion of 3D atomic coordinates with chemical valence, charge distribution, and magnetic fields, paving the way for truly high-dimensional, multi-modal quantitative analysis.

Material service responses—from millisecond-scale tensile fracture to nanosecond-scale phase transitions and picosecond-scale charge density wave (CDW) transitions—are intimately linked to the dynamic evolution of their microstructure over time. To deeply understand structure-property relationships, multi-timescale imaging capabilities covering milliseconds, microseconds, nanoseconds, and down to femtoseconds are needed, with the ultimate goal of combining femtosecond temporal resolution with atomic spatial resolution to directly observe electron dynamics and lattice vibrations [31]. The challenge for time-resolved TEM lies in maintaining high spatial resolution while achieving high temporal resolution and effectively coupling observations across different time dimensions. Current approaches can be divided into two main categories. Continuous imaging from seconds to microseconds for non-repeatable phenomena relies on fast detectors, and when combined with deep learning denoising, spatial resolution can reach the atomic scale. For instance, Crozier et al [25]. used convolutional neural network denoising to continuously observe the dynamic evolution of metastable atoms on Pt nanoparticle surfaces at microsecond timescales. Another approach is the pump-probe mode driven by ultrafast lasers [32]. This mode achieves higher temporal resolution from microseconds to femtoseconds but, because it relies on signal accumulation over many events, is primarily applicable to reversible processes induced by photoexcitation. For example, Deng et al. studied InP decorated with Ag nanoparticles at femtosecond-nanometer scales, finding that the nanoparticles significantly enhanced light absorption and carrier excitation via localized surface plasmon resonance, and greatly accelerated charge diffusion while extending carrier lifetimes [33]. Sun et al. and Danz et al. investigated photoinduced CDW phase transitions in 1T-TaS_2-ₓSeₓ and TaS₂ systems on femtosecond to picosecond timescales [34,35]. Furthermore, initial attempts to combine 3D imaging with time-resolved imaging have also been demonstrated by various groups [36,37].

These spatiotemporal extensions pose significant challenges to signal acquisition efficiency—extremely short exposure times and the massive data required for 3D reconstruction make insufficient signal-to-noise ratio a key bottleneck limiting the advancement of spatiotemporal imaging towards greater Material Adaptiveness. Developing bright, highly coherent electron sources is a potential pathway to achieving femtosecond temporal resolution while maintaining sub-angstrom spatial resolution in TEM, for example, developing coherent ultrafast photoemission sources based on single quantized states in carbon nanotubes [38]. Additionally, reducing the required imaging area through sparse imaging and compressed sensing reconstruction algorithms is another important way to enhance temporal resolution [39]. Clearly, in the near future, TEM will progressively evolve towards synchronously acquiring and fusing characteristic signals spanning different spatiotemporal scales, thereby pushing the characterization of multi-dimensional information—chemical composition, electronic structure, local electric and magnetic fields—to new heights.

2.1.3 Complex environment imaging

The actual service conditions for materials are often extremely complex, differing vastly from the high-vacuum, room-temperature environment of conventional TEM. For example, aerospace alloys must withstand high temperatures, battery materials face complex solid-liquid interfaces, and zeolite catalysts in coal chemical engineering operate in gas-liquid-solid multiphase environments. Developing Material Adaptiveness in TEM aims to expand its capability to characterize material behavior under these realistic conditions. In-situ/operando TEM techniques strive to simulate or replicate real operating conditions inside the microscope by developing environmental holders with chip-based encapsulation or by using differential pumping to create a low-vacuum gaseous environment directly within the column [40]. Currently, in-situ TEM techniques incorporating single stimuli or fields (e.g., mechanical, thermal, optical, electrical, gaseous, liquid, magnetic) are relatively mature, and systems capable of applying coupled multi-environmental fields are emerging. For instance, Ling et al. systematically studied the structural response of MIL-53 metal−organic frameworks (MOFs) under different environmental fields using various cryogenic, liquid, and gas holders [41]. He et al., combining in-situ mechanical TEM with atomic force microscopy, observed in real-time the formation of metastable ordered atomic layers in metallic asperities during friction [42]. Numerous reviews also provide excellent summaries of in-situ TEM techniques for different material systems [43–45].

However, introducing complex environments brings new challenges: reduced imaging resolution and increased operational complexity, hindering their wider application. Addressing this requires optimizing holder design at the hardware level to enhance system robustness and minimize environmental interference. At the software level, adaptive control feedback algorithms are needed to increase automation, and machine learning-based video recognition can assist researchers in automatically identifying characteristic structural evolution from massive dynamic datasets. When these synergistic hardware and software breakthroughs are achieved, in-situ/operando TEM will further help researchers establish direct causal links between microstructural evolution and macroscopic performance under near-realistic service conditions, thereby expanding the boundaries of TEM Material Adaptiveness from static, isolated observations to dynamic, coupled operational environments.

2.2 Knowledge adaptiveness: empowering a broader community of researchers

As Material Adaptiveness expands TEM's application boundaries, a growing number of researchers from physics, chemistry, biology, geology, and engineering will have urgent needs to characterize their materials at the atomic scale. However, operating an advanced TEM and interpreting complex data requires years of specialized training. Knowledge Adaptiveness aims to encapsulate, transfer, and empower through intelligence and automation, facilitating efficient cross-disciplinary utilization of TEM and enabling breakthroughs. It encompasses development on three levels:

2.2.1 Expert knowledge transfer

Using AI for knowledge distillation and transfer learning, expert system models can be built [46]. These models encode complex instrument calibration procedures (e.g., optic alignment, focusing, drift correction) and data analysis experience into software modules. Users, without needing deep understanding of the underlying physics or operational details, can select predefined “material type - scientific question” combinations via a user-friendly interface. The system then automatically executes the entire workflow—from sample navigation and imaging condition optimization to data acquisition and preliminary analysis. This “one-click imaging” mode, encapsulating expert knowledge, will significantly lower the barrier for interdisciplinary researchers. For example, Lebas et al. developed a customized holder and automated 3D reconstruction software enabling rapid continuous tilt-rotation data acquisition under environmental conditions, simplifying sample preparation and characterization while reducing electron dose [47]. For automating parameter selection in complex techniques like ptychography, which currently relies heavily on expert intuition, automatic parameter optimization strategies like Bayesian optimization and large language models show great potential for future fully automated operation [48,49]. Wang et al. applied variational autoencoder deep learning to in-situ TEM data analysis, successfully decoding lattice motion information from complex datasets of nanocrystals during annealing [50].

Currently, automated TEM analysis encapsulated with expert knowledge is still in its early stages, with most efforts being small-scale attempts by specialized microscopy groups on a limited set of materials. In the future, close collaboration between experts from different materials domains and electron microscopists will be essential to develop specialized automated TEM analysis packages tailored to various materials and specific scientific questions. For example, in the field of low-dimensional materials, considerable research has already been devoted to the automated identification and classification of various defects [51]. For nanoparticles, significant progress has also been made in particle counting, size measurement, and morphological analysis [52]. Moreover, these single-tool analysis packages are designed for different instruments, materials, and scientific problems. They will need to adopt uniform or mutually compatible interfaces. This also calls for the scientific community to build infrastructure-level facilities that standardize data and analytical methods. At the same time, developing more intuitive and user-friendly graphical user interfaces (GUIs) is an important pathway to broaden the adoption of these automated analysis tools. With the rapid advancement of generative artificial intelligence, automated TEM analysis can be expected to mature and become widespread in the near future. For instance, in a recent study, Chen et al. developed a large language model driven multi-agent platform called EMSeek, which automatically performs the entire analysis workflow from electron microscopy images to atomic structures, material properties, and literature validation, compressing weeks of manual work into just a few minutes [53].

In the future, combining pre-trained expert analysis modules with encapsulated, easy-to-call advanced imaging methods could further promote interdisciplinary integration, for instance, extending the analysis of nanogram-level lunar samples to microgram or even milligram levels, thereby more deeply revealing the Moon's evolution [54,55].

2.2.2 Intelligent automation

Automation is fundamental to enhancing instrument utilization. By integrating machine vision, robotic control, and AI decision-making, processes such as automatic holder loading, sample positioning, imaging condition optimization, and long-duration or large-dataset acquisition can be automated. While fully automated operation might lack the flexibility of human experts for extremely complex samples, its advantages lie in “trading time for quality” and “trading quantity for discovery.” It can generate massive standardized datasets, automatically filter qualified data frames using built-in quality assessment algorithms, vastly accelerating the discovery of statistical laws and serving machine learning training. In structural biology and cryo-EM, the massive datasets of protein images rely heavily on automated acquisition routines. Based on this, more specialized automated processing software has been developed for steps like image alignment, particle picking, and structure reconstruction [56]. The rapid progress in AI and the continuous growth of specialized datasets is driving deeper TEM automation. Using expertly curated data and pre-trained models, a closed-loop workflow integrating acquisition, recognition, and quality assessment can be built [57,58]. If this workflow can be further integrated with developments in Material Adaptiveness, it could in the near future to add new spatial, temporal, environmental, and information dimensions to fully automated TEM characterization for industrial applications. For instance, in fields like the semiconductor industry requiring high-throughput, high-precision defect inspection, a Knowledge-Adaptable TEM system could be designed with a hierarchical workflow. A screening module, using automated high-speed TEM and image recognition, rapidly scans numerous similar samples to pinpoint potential defects or anomalous regions. Subsequently, an expert analysis module automatically switches key identified regions to higher-resolution imaging modes (like ptychography) or triggers more in-depth analysis (like atomic-resolution EDX mapping), or even allows remote expert intervention for complex operations. Currently, fully automated inspection in the semiconductor industry is largely confined to routine tasks like elemental analysis and geometric defect detection. However, emerging techniques like multi-slice ptychography already achieve atomic-scale imaging of 3D buried devices like gate-all-around transistors with sub-angstrom lateral and nanometer depth resolution, enabling quantitative measurement of interface roughness, strain, and defect distributions [59]. Combining the “broad screening” of routine rapid inspection with the “precise analysis” of key regions will further refine the industry's demand for high-throughput, high-reliability characterization [60,61].

Beyond industrial defect inspection, intelligent automation also promises to bridge the gap between atomic scale observations and macroscopic material properties. By combining automated large area image acquisition with machine learning based analysis, TEM characterization can be extended from nanoscale regions to micrometer or even millimeter scale areas. Looking forward, we envision that high throughput atomic resolution TEM, analogous to the high throughput scanning electron microscope study by Ju et al. on nickel base superalloys, could identify every atomic column across areas of tens of square micrometers or more, enabling statistically representative quantification of defects, interfaces, and precipitates that directly correlate with bulk performance [62]. As a step in this direction, Cho et al. employed deep learning assisted high throughput TEM to statistically characterize over 440,000 Co₃O₄ nanocrystals, revealing size resolved shape evolution and the onset radius for growth regime transitions [63]. These examples illustrate that intelligent automation, when integrated with high throughput imaging and machine learning, can elevate TEM from probing localized atomic structures to delivering statistically representative, multiscale structural information that directly informs macroscopic material behavior.

2.2.3 Cloud operation

The development of Material and Knowledge Adaptiveness will inevitably stimulate an explosion in demand for advanced TEM characterization. However, top-tier instruments are expensive, maintenance-intensive, and geographically unevenly distributed, creating a contradiction between their scarcity and rapidly growing needs. Through cloud-based and grid-based scheduling, an efficient, equitable, and collaborative intelligent TEM resource network can be constructed. “Cloudification” can offload complex computational tasks like image processing, 3D reconstruction, and big data analysis to cloud servers, allowing users to submit tasks and retrieve results via lightweight clients without needing local high-performance computing clusters. For example, Zuo et al.'s cloud EMAPS microscopy simulation platform overcomes limitations of traditional desktop software by offering on-demand access, data sharing, and high performance computing (HPC) capabilities, providing a more convenient and efficient solution for interactive simulation of complex processes like electron diffraction and aberration correction optics [64]. Going further, regional or national “intelligent TEM grids” could be established, comprising a central intelligent scheduling hub and multiple distributed, highly automated TEM terminals. Researchers submit samples online along with scientific questions and characterization requirements. The scheduling hub optimizes task allocation based on sample properties, required techniques, each terminal's expertise, and real-time load. Samples are physically shipped to the designated terminal, characterized in the automated system, and data is automatically uploaded to the cloud for analysis and report generation. This model could enable “next-day validation of research ideas": an idea generated and sample prepared during the day could be automatically tested overnight on the most suitable microscope, with preliminary results available the next morning. This would dramatically accelerate the iteration cycle of scientific discovery, fostering a new paradigm of “agile materials R&D.” As multi-dimensional, cross-disciplinary, high-throughput TEM scientific data accumulates exponentially, such data, after project completion and appropriate anonymization and standardization, could be curated into open, traceable scientific databases. This would critically address a core bottleneck for generative AI development—the scarcity of high-quality domain-specific data. A training repository built upon big TEM data would become a cornerstone for developing next-generation materials science AI (for property prediction, failure analysis, structural optimization, etc.), effectively mitigating the issue of data scarcity in specialized fields.

3 Advanced TEM Applications and Prospects in Transformative Material Systems

The TEM Adaptiveness Revolution directly addresses the need to characterize structure-property relationships in diverse, transformative materials under realistic conditions. Fields such as advanced functional ceramics, high-strength aerospace materials, and multi-purpose energy catalytic materials all grapple with challenges related to functional unit response, structural evolution, and active site reconstruction, necessitating TEM's multi-field coupling imaging, spatiotemporal multi-dimensional deciphering capabilities, and intelligent tools. This section will exemplify current advanced TEM applications in these three material classes and outlook how the TEM Adaptiveness Revolution can empower key breakthroughs in transformative materials.

3.1 Advanced functional ceramics

The macroscopic properties of advanced functional ceramics arise from the coupled evolution of multi-dimensional order parameters—lattice, charge, orbital, spin—under external fields. Understanding atomic-scale mechanisms like domain wall dynamics and interfacial thermal transport is key to designing next-generation ferroelectric, piezoelectric, and multiferroic devices.

For example, in ZrO₂ ferroelectric thin films, researchers using multi-slice ptychography directly observed 1D charged domain walls in free-standing films for the first time, with both width and thickness confined to a single sub-unit cell (Fig. 3 A i−ii) [65]. By quantitatively analyzing the ptychographic phase information, they precisely determined the occupancy of individual oxygen atomic columns, revealing a unique screening mechanism where head-to-head walls compensate bound charge through excess oxygen ions, and tail-to-tail walls via oxygen vacancies. Combining integrated differential phase contrast imaging (Fig. 3 A iii), they further observed the in-situ dynamic motion of these 1D charged walls, finding that the motion is essentially an electric-field-driven migration of oxygen ions, thus revealing a strong coupling between ferroelectric switching and oxygen ion transport.

Interfacial thermal transport is a critical scientific issue for further semiconductor miniaturization. In a study of AlN-SiC heterointerfaces, researchers used an in-situ heating holder to establish a stable temperature gradient and, combined with vibrational electron energy-loss spectroscopy, achieved sub-nanometer imaging of non-equilibrium phonon populations near the interface for the first time (Fig. 3 B i−iii) [66]. They found the temperature drop occurred within a narrow ~2 nm region and that the occupation of interfacial phonon modes showed significant asymmetry under forward and reverse heat flow (Fig. 3 B iv−v). Coupled with molecular dynamics simulations, the study revealed the microscopic mechanism where interfacial phonons preferentially couple with high-energy bulk phonons via inelastic scattering, providing direct experimental evidence for energy transport across buried interfaces.

These studies show that advanced TEM has evolved from pure structural imaging to the comprehensive characterization of multi-dimensional functional units like domain wall and phonons. Looking ahead, enhanced Material Adaptiveness will push complex environment imaging towards coupled multi-fields (electric, stress, temperature), e.g., real-time tracking of domain wall motion and local strain correlation in piezoelectrics. Full-information imaging will further integrate ptychography's high spatial resolution with the chemical state information provided by EELS. Knowledge Adaptiveness, via deep learning-based automatic recognition algorithms, will transform domain wall statistical analysis into standardized tools, and combined with automated high-throughput characterization and remote intelligent scheduling, will build a new paradigm for agile ceramics R&D.

3.2 High-strength structural materials

The mechanical behavior of high-strength structural materials (e.g., superalloys, ceramic matrix composites) under extreme service conditions—such as dislocation motion, grain boundary sliding, and phase transformation evolution—directly determines component lifetime and safety. Atomic-scale dynamic observation is crucial for unveiling their deformation and failure mechanisms.

The intrinsic brittleness of ceramics has long been a core challenge limiting their use as structural materials. In research on toughening ceramics by “borrowing dislocations,” researchers used dark-field TEM to directly observe the atomic configuration of an ordered bonding interface between La₂O₃ ceramic and Mo metal [67]. This engineered heterogeneous interface, characterized by high coherency and strong chemical bonding, provided a channel for dislocation transmission from the metal into the ceramic. Using in-situ tensile TEM, they tracked this dynamic process in real time. The observations revealed dislocation bundles originating in the Mo metal, crossing the heterointerface, and continuously sliding and multiplying within the La₂O₃ ceramic (Fig. 4 A i−vi). This unprecedented dislocation activity enabled the intrinsically brittle ceramic to exhibit up to 39.9% tensile ductility at room temperature (Fig. 4 A vii−ix). This work not only breaks the intrinsic brittleness limitation of ceramics but also showcases the immense potential of atomic-resolution complex environment imaging in solving key materials science problems.

For metals, grain boundary behavior governs plastic deformation mechanisms. In a study of grain boundary sliding in Pt bicrystals, researchers used in-situ atomic-resolution TEM to track for the first time the sliding process of a general tilt grain boundary at room temperature in real-time (Fig. 4 B i−ii) [68]. Using automated atomic column tracking and atomic-scale strain analysis, they discovered two modes of boundary sliding: direct atomic sliding along the boundary, and sliding accompanied by atomic plane transfer—the latter occurring via a five-membered ring core unit of a grain boundary Lomer lock, enabling atomic column transfer across the interface and transport along it, leading to a change in the number of lattice planes on one side of the boundary (Fig. 4 B iii). This finding provides direct experimental evidence for understanding grain boundary-mediated plasticity in nanocrystalline metals and highlights the synergistic value of spatiotemporal multi-dimensional imaging and automated tracking algorithms in studying complex mechanical behavior.

Looking ahead, enhanced Material Adaptiveness will push complex environment imaging towards coupled mechanical-thermal-atmospheric fields. For example, simultaneously applying thermal gradients and tensile stress in Ni-based single crystal superalloys to track in-situ the coupled mechanisms of dislocation climb, grain boundary sliding, and void evolution. Full-information imaging will fuse the high spatial resolution of 3D reconstruction with quantitative chemical information from EDX/EELS, enabling simultaneous deciphering of the 3D atomic structure of nanoprecipitates and local chemical segregation at interfaces in Re-doped superalloys. Combining AI-driven trajectory tracking algorithms, the diffusion paths and segregation behavior of Re atoms could be traced at the atomic scale. In the Knowledge Adaptiveness domain, for dislocation-precipitate interactions, deep learning-based automatic recognition algorithms can be developed to rapidly study dislocation depinning frequencies from massive in-situ videos. Automated operation and hierarchical screening will support Materials Genome Engineering initiatives, enabling high-throughput mechanical testing and structural characterization. Remote operation and intelligent scheduling will transform scarce aberration-corrected TEM resources into a network for aerospace materials characterization, enabling multi-scale prediction from microscopic mechanisms to macroscopic properties.

3.3 Multi-purpose energy catalytic materials

The performance of energy catalytic materials critically depends on atomic-scale dynamic processes at solid-liquid and solid-gas interfaces, including active site reconstruction, reaction intermediate evolution, and product nucleation and growth. Conventional ensemble-averaged characterization techniques are inadequate for capturing these transient behaviors, creating an urgent need for in-situ dynamic observation techniques.

For example, the liquid-solid conversion mechanism of polysulfides in lithium-sulfur batteries has long been debated. Researchers utilized in-situ liquid phase electrochemical TEM to track this process in real-time at the nanoscale [69]. By constructing an electrochemical liquid cell (Fig. 5 A i−ii), they observed an active center-induced aggregation effect: soluble polysulfides enriched near Mo nanoclusters, forming droplet-like dense phases, which then triggered collective charge transfer, leading to the instantaneous deposition of non-equilibrium Li₂S nanocrystals (Fig. 5 A iii−iv). In the absence of active centers, the reaction followed the traditional stepwise single-molecule pathway, first forming a Li₂S₂ intermediate phase before converting to Li₂S. This work not only clarified the nucleation and growth mechanism of Li₂S but also underscored the core value of liquid-phase electrochemical environments and multi-signal fusion in catalysis research.

In the field of gas-phase catalytic material synthesis, the chemical vapor deposition process of MoS₂ contains rich dynamic mechanisms. Researchers used in-situ environmental TEM to create a chemical vapor deposition (CVD) growth environment within a micro-reactor cell, tracking the complete pathway from nucleation to early growth of MoS₂ at the atomic scale (Fig. 5 B i−ii) [70]. The experiments revealed that MoS₂ does not directly form crystalline nuclei (Fig. 5 B iii−iv). Instead, it first undergoes an amorphous cluster stage, transforms into a 2D amorphous layered embryo, and only undergoes in-plane ordering to become a crystalline nucleus after exceeding a critical size. Furthermore, they captured the process of nuclei merging and growing through oriented attachment. This study provided the first experimental confirmation of the multi-step nucleation mechanism for 2D materials, offering key experimental evidence for controlled synthesis.

These studies demonstrate that the deep integration of operational environments with high spatiotemporal resolution imaging is pushing catalysis research from indirect inference towards direct observation. Looking ahead, enhanced Material Adaptiveness will push complex environment imaging towards coupled mechanical-thermal-atmospheric-electrochemical fields. For instance, simultaneously applying thermal gradients and atmosphere changes in solid oxide fuel cells to track in-situ the oxygen ion transport pathways at the three-phase boundary. A key goal is spatiotemporal multi-dimensional imaging that combines femtosecond temporal resolution with 3D atomic resolution. Achieving this would enable direct observation of transient intermediates in catalytic reactions, revealing ultrafast processes like photoinduced charge separation and surface adsorption. Complementing this, full-information imaging could simultaneously probe the atomic configuration and electronic state of catalyst active sites within the same nanoscale region. On the Knowledge Adaptiveness front, deep learning-based automatic recognition algorithms can rapidly statistically analyze the dynamic evolution of active sites from massive in-situ videos, and automated operation with hierarchical screening can support high-throughput catalyst screening. As Material Adaptiveness and Knowledge Adaptiveness develop synergistically, catalysis research will transition from inferring mechanisms based on ensemble-averaged information of steady states to directly observing processes via dynamic imaging, propelling catalyst design from empirical trial-and-error towards rational construction.

4 Conclusion and Outlook

In the current wave of developing new productive forces, the Adaptiveness Revolution in transmission electron microscopy is dedicated to deploying this powerful microscope to every key node in the exploration of the real, dynamic, and complex material world, ensuring that every researcher in need can use it conveniently and effectively. This paradigm shift will have profound impacts: At the fundamental science level, it will extend atomic-scale studies to encompass lattice, charge, orbital, and spin degrees of freedom in numerous materials, as well as their dynamic behavior in complex environments, fostering new insights in materials science. At the technological R&D level, it will significantly accelerate the development cycle of transformative materials—such as novel semiconductors, fatigue-resistant alloys, high-performance batteries, and efficient catalysts—through agile R&D models and data-driven discovery. At the industrial application level, it will provide tools for atomic-scale quality control and R&D in semiconductor manufacturing, aerospace materials testing, chemical process optimization, and drug structure determination. Ultimately, this series of transformations will deepen our understanding of the material world and empower us to address global challenges in energy, environment, health, and information.

References

[1] Rose H. History of direct aberration correction. The Netherlands: Academic Press, 2009

[2] Chen Z , Jiang Y , Shao Y T . et al. Electron ptychography achieves atomic-resolution limits set by lattice vibrations. Science, 2021, 372(6544): 826–831

[3] Yang W , Sha H , Cui J . et al. Local-orbital ptychography for ultrahigh-resolution imaging. Nature Nanotechnol, 2024, 19(5): 612–617

[4] Zhang D , Zhu Y , Liu L . et al. Atomic-resolution transmission electron microscopy of electron beam–sensitive crystalline materials. Science, 2018, 359(6376): 675–679

[5] Gao A , Gu L . Insight into materials science from a reductionist perspective. Innov Mater, 2023, 1(1): 100009

[6] Williams D B, Carter C B. Transmission electron microscopy: A textbook for materials science. Springer New York, NY, 2009

[7] Susi T , Meyer J C , Kotakoski J . Quantifying transmission electron microscopy irradiation effects using two-dimensional materials. Nature Rev Phys, 2019, 1(6): 397–405

[8] Philipp H T , Tate M W , Shanks K S . et al. Very-high dynamic range, 10, 000 frames/second pixel array detector for electron microscopy. Microsc Microanal, 2022, 28(2): 425–440

[9] Zambon P , Vávra J , Montemurro G . et al. High-frame rate and high-count rate hybrid pixel detector for 4D STEM applications. Front Phys, 2023, 11: 1308321

[10] Li X , Mooney P , Zheng S . et al. Electron counting and beam-induced motion correction enable near-atomic-resolution single-particle cryo-EM. Nature Methods, 2013, 10(6): 584–590

[11] Zhu Y , Ciston J , Zheng B . et al. Unravelling surface and interfacial structures of a metal–organic framework by transmission electron microscopy. Nature Mater, 2017, 16(5): 532–536

[12] Niculae A , Bornschlegl M , Eckhardt R . et al. Large solid angle silicon drift detectors for edx analysis in TEM. Microsc Microanal, 2014, 20(S3): 1124–1125

[13] Ding Q , Zhang Y , Chen X . et al. Tuning element distribution, structure and properties by composition in high-entropy alloys. Nature, 2019, 574(7777): 223–227

[14] Brown H G , Smith D , Wardle B C . et al. Square condenser apertures for square cameras in low-dose transmission electron microscopy. Nature Methods, 2024, 21(4): 566–568

[15] Song B , Ding Z , Allen C S . et al. Hollow electron ptychographic diffractive imaging. Phys Rev Lett, 2018, 121(14): 146101

[16] Schwartz J , Di Z W , Jiang Y . et al. Imaging 3D chemistry at 1 nm resolution with fused multi-modal electron tomography. Nature Commun, 2024, 15(1): 3555

[17] Inada H , Su D , Egerton R F . et al. Atomic imaging using secondary electrons in a scanning transmission electron microscope: experimental observations and possible mechanisms. Ultramicroscopy, 2011, 111(7): 865–876

[18] Dyck O , Almutlaq J , Lingerfelt D . et al. Direct imaging of electron density with a scanning transmission electron microscope. Nature Commun, 2023, 14(1): 7550

[19] Bing Mo , Zhuang Guo , Yang Li . et al. In situ investigation of the valence states of iron-bearing phases in chang’E-5 lunar soil using FIB, AES, and TEM-EELS techniques. Atom Spectrosc, 2022, 43(1): 1–7

[20] Wang Z , Zou C , Lin C . et al. Correlating the charge-transfer gap to the maximum transition temperature in Bi₂Sr₂Ca_n-1Cu_nO_2n+4+δ. Science, 2023, 381(6654): 227–231

[21] Dong Z , Huo M , Li J . et al. Visualization of oxygen vacancies and self-doped ligand holes in La₃Ni₂O_7-δ. Nature, 2024, 630(8018): 847–852

[22] Seki T , Ikuhara Y , Shibata N . Theoretical framework of statistical noise in scanning transmission electron microscopy. Ultramicroscopy, 2018, 193: 118–125

[23] Thornley W , Sullivan-Allsop S , Cai R . et al. Noise2Void for denoising atomic resolution scanning transmission electron microscopy images. npj Comput Mater, 2026, 12(1): 68–81

[24] Sadri A , Petersen T C , Terzoudis-Lumsden E W C . et al. Unsupervised deep denoising for four-dimensional scanning transmission electron microscopy. npj Comput Mater, 2024, 10(1): 243

[25] Crozier P A , Leibovich M , Haluai P . et al. Visualizing nanoparticle surface dynamics and instabilities enabled by deep denoising. Science, 2025, 387(6737): 949–954

[26] Luo Z , Feng M , Gao Z . et al. Exploring structure diversity in atomic resolution microscopy with graph. Adv Mater, 2025, 37(15): 2417478

[27] Yuan W , Yao B , Tan S . et al. Generative learning of morphological and contrast heterogeneities for self-supervised electron micrograph segmentation. npj Comput Mater, 2025, 11(1): 322–336

[28] Xiong H , Liu Z , Chen X . et al. In situ imaging of the sorption-induced subcell topological flexibility of a rigid zeolite framework. Science, 2022, 376(6592): 491–496

[29] Miao J , Ercius P , Billinge S J L . Atomic electron tomography: 3D structures without crystals. Science, 2016, 353(6306): aaf2157

[30] Moniri S , Yang Y , Ding J . et al. Three-dimensional atomic structure and local chemical order of medium- and high-entropy nanoalloys. Nature, 2023, 624(7992): 564–569

[31] Alcorn F M , Jain P K , Van Der Veen R M . Time-resolved transmission electron microscopy for nanoscale chemical dynamics. Nat Rev Chem, 2023, 7(4): 256–272

[32] LaGrange T , Cattaneo P , Barwick B . et al. Laser-driven ultrafast transmission electron microscopy. Nat Rev Methods Primers, 2025, 5(1): 61–82

[33] Deng S , Qiu M , Chen X . et al. Imaging silver-nanoparticle-decoration-boosted charge separation and transport in inp by scanning ultrafast electron microscopy. Phys Rev Mater, 2026, 10(2): 24606

[34] Sun K , Sun S , Zhu C . et al. Hidden CDW states and insulator-to-metal transition after a pulsed femtosecond laser excitation in layered chalcogenide 1T-TaS _{2- x}Se_x. Sci Adv, 2018, 4(7): eaas9660

[35] Danz T , Domröse T , Ropers C . Ultrafast nanoimaging of the order parameter in a structural phase transition. Science, 2021, 371(6527): 371–374

[36] Kwon O H , Zewail A H . 4D electron tomography. Science, 2010, 328(5986): 1668–1673

[37] Meana-Pañeda R , Ji C , Van C T S . et al. Time-resolved atomic-resolution brownian tomography of single nanocrystals reveals size-dependent dynamics. Sci Adv, 2025, 11(32): eady1413

[38] Li C , Guan M , Hong H . et al. Coherent ultrafast photoemission from a single quantized state of a one-dimensional emitter. Sci Adv, 2023, 9(41): eadf4170

[39] Robinson A W , Wells J , Moshtaghpour A . et al. Real-time four-dimensional scanning transmission electron microscopy through sparse sampling. Chin Phys B, 2024, 33(11): 116804

[40] Singh R. In-situ tem: principle, working, application and challenges. Springer Nature Switzerland, 2024

[41] Wei L , Sun T , Shi Z . et al. Guest-adaptive molecular sensing in a dynamic 3D covalent organic framework. Nat Commun, 2022, 13(1): 7936

[42] He Y , She D , Liu Z . et al. Atomistic observation on diffusion-mediated friction between single-asperity contacts. Nat Mater, 2022, 21(2): 173–180

[43] Zhou Z , Xiaobo Chen , Dongxiang Wu . et al. In situ electron microscopy: Atomic-scale dynamics of metal oxidation and corrosion. npj Mater Degrad, 2025, 9: 28

[44] Chao H Y , Venkatraman K , Moniri S . et al. In situ and emerging transmission electron microscopy for catalysis research. Chem Rev, 2023, 123(13): 8347–8394

[45] Han Y , Wang L , Cao K . et al. In situ TEM characterization and modulation for phase engineering of nanomaterials. Chem Rev, 2023, 123(24): 14119–14184

[46] Gou J , Yu B , Maybank S J . et al. Knowledge distillation: A survey. Int J Comput Vis, 2021, 129(6): 1789–1819

[47] Lebas L M , Masenelli-Varlot K , Trillaud V . et al. Fast automatic multiscale electron tomography for sensitive materials under environmental conditions. Commun Eng, 2025, 4(1): 149–159

[48] Cao M C , Chen Z , Jiang Y . et al. Automatic parameter selection for electron ptychography via bayesian optimization. Sci Rep, 2022, 12(1): 12284

[49] Yin X , Shi C , Deng J . et al. PEAR: A knowledge-guided autonomous pipeline for ptychography enabled by large language models. Microsc Microanal, 2024, 30(Supplement_1): ozae044.184

[50] Wang X , Yan C , Ondry J C . et al. An artificial intelligence’s interpretation of complex high-resolution in situ transmission electron microscopy data. Matter, 2024, 7(1): 175–190

[51] Kalinin S V , Mukherjee D , Roccapriore K . et al. Machine learning for automated experimentation in scanning transmission electron microscopy. npj Comput Mater, 2023, 9(1): 227–242

[52] Aviles K M , Lear B J . Practical guide to automated tem image analysis for increased accuracy and precision in the measurement of particle size and morphology. ACS Nanosci Au, 2025, 5(3): 117–127

[53] Chen G, Yuan W, You F. Bridging electron microscopy and materials analysis with an autonomous agentic platform. Science Advances, 2026, 12(eaed0583)

[54] Su B , Chen Y , Wang Z . et al. South pole–aitken massive impact 4. 25 billion years ago revealed by Chang’e-6 samples. Natl Sci Rev, 2025, 12(6): nwaf103

[55] Li J , Xing L , Gong Z . et al. Magnetic signatures and origins of ferromagnetic minerals in chang’e-6 lunar farside soils. Nature Commun, 2025, 16(1): 6218

[56] Mendez J H , Chua E Y D , Paraan M . et al. Automated pipelines for rapid evaluation during cryoem data acquisition. Curr Opin Struct Biol, 2023, 83: 102729

[57] Chen Q , Xu Z , Dai H . et al. A large-scale curated and filterable dataset for cryo-EM foundation model pre-training. Sci Data, 2025, 12(1): 961–990

[58] Yan Y , Fan S , Yuan F . et al. A comprehensive foundation model for cryo-EM image processing. Nat Methods, 2026, 23(1): 88–95

[59] Karapetyan S, Zeltmann SE, Wilk G, et al. 3D atomic-scale metrology of strain relaxation and roughness in gate-all-around transistors via electron ptychography. Nature Commun, 2026

[60] Flanagan J , Sixtos A , Hakala C . et al. Smart automation: machine learning enabled workflow for logic and dram. Microsc Microanal, 2021, 27(S1): 61–61

[61] Strauss M, Li C, Hakala C, et al. Automated (S)TEM metrology characterization of gate-all-around and 3D NAND devices. In: Metrology, Inspection, and Process Control XXXVII, 2023, 1249639

[62] Ju Y , Li S , Yuan X . et al. A macro-nano-atomic–scale high-throughput approach for material research. Sci Adv, 2021, 7(49): eabj8804

[63] Cho M G , Sytwu K , Rangel DaCosta L . et al. Size-resolved shape evolution in inorganic nanocrystals captured via high-throughput deep learning-driven statistical characterization. ACS Nano, 2024, 18(43): 29736–29747

[64] Zuo J M , Zhu X R , Ang E . et al. CloudEMAPS: A cloud computing environment for electron microscopy application simulations. Microscopy Today, 2021, 29(1): 24–27

[65] Zhong H , Wang S , Zhang Q . et al. Observation of one-dimensional, charged domain walls in ferroelectric ZrO₂. Science, 2026, 391(6783): 407–411

[66] Liu F , Mao R , Liu Z . et al. Probing phonon transport dynamics across an interface by electron microscopy. Nature, 2025, 642(8069): 941–946

[67] Dong L R , Zhang J , Li Y Z . et al. Borrowed dislocations for ductility in ceramics. Science, 2024, 385(6707): 422–427

[68] Wang L , Zhang Y , Zeng Z . et al. Tracking the sliding of grain boundaries at the atomic scale. Science, 2022, 375(6586): 1261–1265

[69] Zhou S , Shi J , Liu S . et al. Visualizing interfacial collective reaction behaviour of Li–S batteries. Nature, 2023, 621(7977): 75–81

[70] Ye H , Wu C , Cao D . et al. Atomically resolved two-dimensional amorphous nuclei formed during MoS₂ chemical vapor deposition. Science, 2026, 391(6785): 622–627

RIGHTS & PERMISSIONS

The Author(s). This article is published by Higher Education Press.