Pseudotime uncovers restructuring dynamics and mitigators of metal catalyst deactivation Authors: Wenhao Yuan, Zhilong Wang, Fengqi You Published: 2026 ABSTRACT Structural reconstruction is a major cause of heterogeneous catalyst deactivation, yet elucidating the restructuring mechanism remains difficult due to underresolved dynamics. Here, by introducing pseudotime, we recover dynamics and identify the key determinant of nanoparticle sintering from ex situ electron microscopy alone, beyond the general view that postmortem observations provide only end-state structures. Pseudotemporal ordering of SiNx-supported gold across asynchronous aging states reveals that interparticle distance controls stability against migration and coalescence, with a well-defined threshold quantified. In situ gas cell tracking validated this threshold with only a 2-angstrom deviation and confirmed markedly suppressed particle mobility beyond it, allowing sintering prediction with 92% accuracy, thereby establishing proximity as a quantitative handle for improving sinter resistance. The proximity effect resolved here illustrates pseudotime as a new lens for understanding time-resolved phenomena in catalysis, opening the possibility of probing the restructuring dynamics beyond in situ or operando characterizations. INTRODUCTION Heterogeneous catalysts underpin more than 90% of chemical production and ∼35% of global GDP (1), yet they inevitably undergo structural reconstruction during operation, often leading to deactivation (2–5). Such restructuring is governed not only by reaction conditions and time but also, more fundamentally, by intrinsic catalyst properties. Understanding this property dependence is central to the rational design of stable catalysts. Sintering of supported metal nanocatalysts (6, 7), for example, can be mitigated by either homogenizing chemical potentials [e.g., size or shape uniformity (7)] or strengthening particle/atom anchoring [e.g., metal-support interactions (8–10) or encapsulation/fixation (11–13)]. Despite advances, strategies targeting intraparticle morphology or particle-support interactions have rarely incorporated interparticle proximity, as its role in sintering remains insufficiently resolved. Particle proximity has emerged as a key determinant of catalytic performance, given its ability to modulate activity (14–18), selectivity (19–21), and stability (22–25). However, its quantification is often reduced to simplified estimates [e.g., statistical (26) or geometric (27) models], imposing three constraints. Inference is indirect, derived from metal loading. Idealized particle sizes and spatial distributions are assumed. Ensemble averaging obscures particle-level variability. Together, these factors compromise accuracy for real catalyst systems. Complementing these estimation approaches, atomistic simulations can provide deeper mechanistic insights (28, 29). For example, first-principles–based kinetic Monte Carlo can resolve the interplay between particle migration and coalescence (PMC) and Ostwald ripening (OR) (30), but prescribed configurations and limited sampling constrain population-level quantification of proximity. Electron microscopy with single-particle resolution, in contrast to bulk techniques, allows direct measurements, yet in practice often stops at static comparisons, where properties extracted from ex situ snapshots at selected states (e.g., fresh and aged) are qualitatively correlated with catalytic performance (31–35). Such empirical correlations collapse structural dynamics and hinder resolving how proximity controls restructuring. Notably, a critical parameter, the proximity threshold above which sintering can be substantially suppressed, has yet, to our best knowledge, to be directly determined from the underlying dynamics but is instead abductively inferred from performance correlations (25). Although advances in in situ and operando electron microscopy (3) allow tracking of such processes, their sampling is often constrained by the field of view and thus inherently is a statistically insignificant, or even unrepresentative, fraction of the entire catalyst surface (5, 36). Such sampling bias can further intensify as structural heterogeneity increases (5). This lack of ability to resolve catalyst dynamics calls for an approach to establishing temporal mapping between structural properties and restructuring states from static snapshots captured at a single time point. A similar trade-off between spatial and temporal resolution exists in single-cell sequencing (37), where, like ex situ electron microscopy, individual entities are observed statically at high spatial resolution but cannot be continuously tracked throughout their evolution (38). Pseudotemporal ordering, or pseudotime (39–41), thus leverages the asynchrony of cell development to arrange snapshot-captured entities by progression state along a continuous trajectory, thereby reconstructing underlying dynamics and identifying key regulatory factors. Inspired by this concept, we hypothesize that intrinsic heterogeneity in the structure and local environment of individual catalytic entities (i.e., nanosized metal species) might induce differences in driving forces and thus onset and rates of restructuring, ultimately resulting in asynchronous progression along the aging pathway, as illustrated in Fig. 1. In this sense, pseudotime may exist within ex situ electron microscopy snapshots. Thus, ordering particles along the latent progression could establish a direct mapping between catalyst properties, such as proximity, and sintering dynamics. Concept of PEM. Structurally heterogeneous nanoparticles can exhibit asynchronous progression along the aging pathway, leading to the coexistence of different restructuring states in a single ex situ snapshot. Ordering these particles with asynchronous states along pseudotime can reconstruct the underlying restructuring dynamics. Representative catalyst micrograph patches are cropped from the images in Fig. 5 (A and B) and figs. S63B and S68C for illustrative purposes only. Here, we introduce pseudotemporal electron microscopy (PEM) to uncover the key origins of metal nanocatalyst instability by reconstructing the underlying dynamics. Using SiNx-supported gold nanoparticles as a model system, high-angle annular dark-field scanning transmission electron microscopy (HAADF-STEM) reveals that postreaction ex situ snapshots preserve coexisting asynchronous sintering states across structurally heterogeneous particles. We then identify and quantify the critical proximity threshold governing sintering by tracking the evolution of interparticle distance distributions of these asynchronous restructuring states along the pseudotemporal trajectory. In situ HAADF-STEM tracking validates the PEM-derived threshold, which differs from the in situ observed ground truth by only 2 Å. Particles beyond this proximity exhibit markedly suppressed mobility and hence coalescence, even under severe sintering conditions above the Tammann temperature (TTammann), allowing prediction of sintering events with up to 92% accuracy. Automating PEM with a one-shot (single image–based) segmentation tool further paves the way for transferring across practical catalyst systems under distinct imaging modes. Together, these results establish PEM as a promising approach for linking structural properties to restructuring dynamics from static observations, thereby offering a route to increase temporal resolution in postmortem microscopic characterization. From a catalyst design perspective, the directly determined critical proximity defines a new quantitative descriptor for developing sinter-resistant materials. RESULTS Asynchronicity of catalyst restructuring We began by validating the existence of pseudotime in ex situ electron microscopy observations and elucidating its origin. The model system, consisting of gold nanoparticles supported on the silicon nitride window in a gas cell (Au/SiNx), was subjected to thermal treatment under a reducing H2 atmosphere. Under this setup, both in situ and ex situ HAADF-STEM characterizations were conducted (figs. S1 and S2). Specifically, 200°C was used as the initial temperature, and the temperature was increased in increments of 100° up to 600°C. Each heating interval was performed at a ramp rate of 10°C min−1 and was followed by a 30-min holding period. For the in situ observations, the entire heating and holding sequence was continuously monitored within the same selected region of interest (ROI). This continued tracking for dynamics will be used for validations in the following section. The ex situ observations that constitute the focus of this section were performed by cooling the sample back to 200°C after each holding step (i.e., 300°, 400°, 500°, and 600°C) to capture a corresponding static image set. It should be emphasized that, to ensure that the captured Au nanoparticles had undergone a measurable extent of PMC and could statistically represent the particle population, two experimental specifications were implemented. First, regarding temperature, the TTammann of Au is 396°C (7), above which bulk diffusion becomes observable. Therefore, 400°C was selected as the central aging temperature to ensure the occurrence of PMC sintering. In particular, attention was paid to the temperature range of 300° to 500°C, where particle restructuring dynamics are expected to be most pronounced (will be validated later). Given that the Hüttig temperature (THüttig) of Au is 238°C (7), above which surface atoms start to exhibit mobility, 200°C was chosen as the holding temperature for ex situ imaging as severe sintering, both OR and PMC, is not expected at this state. Second, regarding the ROI, each ROI typically contains limited nanoparticles (<200) to maintain single-particle resolution, such limited sampling may not statistically represent the millions of particles distributed across the entire support surface, as emphasized by Cargnello et al. (36) and Bates (5). To alleviate potential sampling bias, we selected 12 relatively isolated electron irradiation–induced ring-like features (fig. S1) as positional references and maximized their spatial coverage across the heating chip for ex situ observations. Forty-eight ROIs were defined in regions adjacent to these rings (figs. S3 to S5) to avoid analyzing Au nanoparticle distributions altered by electron beam irradiation within the rings. As will be demonstrated later, this spatially distributed ROI choice proved to be essential. To reveal the existence of pseudotime, a representative ROI is presented, whose static ex situ HAADF-STEM snapshots captured after treatments at different temperatures are shown in Fig. 2. At the initial temperature of 200°C in Fig. 2A, spatially proximate particles can form clustering configurations, for instance, the annotated (a) to (d). During subsequent heating (Fig. 2, B to D), some of these configurations [e.g., (a) to (c)] gradually undergo sintering, whereas others [e.g., configuration (d)] remain unchanged throughout the entire aging process. Closer examination of these configurations that eventually sinter reveals a key feature, asynchronicity. That is, even under identical treatment conditions and at the same observation time point, different particles can reside at different stages of the restructuring process toward (or not) sintering. Existence of pseudotime in ex situ electron microscopy. (A to D) Ex situ HAADF-STEM images of a representative region of Au/SiNx recorded initially at 200°C (A) and after sequential 30-min heating treatments at 300°C (B), 400°C (C), and 500°C (D), respectively. Dashed boxes and corresponding enlarged views highlight four multiparticle configurations (a) to (d) exhibiting temporally asynchronous sintering behaviors: configuration (Ca) already sintered in (C), (Cb) undergoing sintering, (Cc) about to sinter, and (Cd) will not sinter. The coexistence of these asynchronous restructuring states illustrates the pseudotemporal nature of spatially dispersed catalysts in ex situ observations. The imaged region here corresponds to region 12-4 in fig. S5, which was selected from region 12 in fig. S1. The scale bar in (A) applies to all main panels (A to D), and the scale bar in (Aa) applies to all enlarged views (Aa to Dd). A direct visualization of such an asynchronous phenomenon is shown in Fig. 2C. After holding at 400°C for 30 min, which is closest to the TTammann threshold of 396°C, ex situ HAADF-STEM simultaneously captures particle configurations undergoing different sintering stages. Specifically, configuration (a) had already completed sintering at this post-400°C window, which was previously observed to be undergoing sintering during the earlier observation at the post-300°C window (Fig. 2B). Configuration (b) was undergoing sintering (Fig. 2C), where two particles exhibited a clear merging morphological shape (the same phenomena are also captured in the representative ROIs in fig. S6), and the completion was observed at the next post-500°C window (Fig. 2D). Notably, PMC events are typically transient and thus usually hard to capture in ex situ. However, as we mentioned, the ex situ imaging was performed after cooling to 200°C. Consequently, it is highly likely that the configuration (b) was at the early stage of coalescence exactly when the temperature began to drop, and the greatly reduced driving force at 200°C prevented completion of coalescence, thereby allowing this intermediate state to be preserved and imaged. In addition to completed and ongoing sintering, configuration (c), which had not yet undergone sintering at this moment (Fig. 2C) but was poised to do so (Fig. 2D), was also observed. This configuration subsequently completed sintering in the next observation window at post-500°C (Fig. 2D). In contrast to configurations (a) to (c), configuration (d) represents a different class of behavior, which did not undergo sintering throughout the entire heating process. Together, these ex situ HAADF-STEM observations across increasing temperatures demonstrate that spatially dispersed particles within a single static image can exhibit asynchronous progression along the temporal pathway of sintering. Such asynchronicity likely originates from the heterogeneity intrinsic in particle structures and configurations, including size, shape, composition, neighbor counts, proximity, and, notably, support-mediated effects including interfacial bonding, charge transfer, particle anchoring, and migration barriers, as they collectively govern the onset and rate of sintering. These properties are often strongly coupled and difficult to disentangle, but it is precisely this heterogeneous coupling that places particles at distinct stages of aging even under identical external treatments (e.g., temperature, gas atmosphere, and electron beam irradiation). As a result, a static ex situ electron micrograph may not merely represent one kinetic state, but rather a snapshot containing multiple temporal states of structural evolution, as exemplified in Fig. 2C and illustrated in Fig. 1. We therefore conclude that for catalysts inherently having structural heterogeneity (e.g., supported nanoparticles here), pseudotime may naturally exist in their ex situ electron microscopy observations. The origin of pseudotime lies in the asynchronous aging behavior induced by their heterogeneity. This observed asynchronicity is structure dependent at the population level, although the occurrence and timing of individual restructuring events remain thermally stochastic (6, 7). Such intrinsic asynchronicity forms the conceptual foundation of PEM proposed in this work. In what follows, structural properties of catalytic entities in PEM will be extracted to link the latent property-restructuring relationships. Critical proximity in pseudotime To apply PEM to catalysts, we adopt the two-step workflow of pseudotemporal ordering (39). The first step, representation, characterizes individual entities (catalytic particles here) by measurable features. The second step, projection (or ordering), places entities along a continuous path that represents the evolution of the process using such features, thereby determining the underlying patterns. Following this scheme, we first performed representation. In electron microscopy characterizations, catalyst structure is typically quantified using particle size distributions (PSD). However, PSD has two intrinsic limitations. First, diameters are measured manually from selected particles, making PSD susceptible to sampling bias and measurement error. Second, PSD captures only single-particle properties. However, catalyst stability depends not only on properties of individual catalysts (e.g., size, shape, and composition) but also on collective properties such as spatial distribution (22). We therefore developed a single image–based deep learning framework, NanoSeg (details will be discussed in the “Pseudotime transferability” section), that segments catalytic entities (representative segmentation results are visualized in figs. S7 to S9) and measures their geometrical metrics such as projected area (as size) and nearest-neighbor distance (NND) (22), enabling quantification of both morphological and dispersion-related properties. The second-step projection is then performed using kernel density estimation (KDE), transforming these discrete properties into continuous probability distributions. The resulting NND distributions quantified from ex situ observation across 48 different ROIs are shown in Fig. 3A and fig. S10, exhibiting two notable features. Critical proximity and size preferences inferred from pseudotime. (A) KDE distributions of the NND of Au/SiNx from 48 ex situ HAADF-STEM regions at different temperature windows. (B) Particle-size KDE distributions for particles with NND below or above the critical proximity threshold. Asterisks indicate statistically significant differences between the two particle-size distributions determined by the Mann-Whitney U test. From top to bottom, the corresponding P values are 3.8 × 10−14, 1.1 × 10−10, 1.2 × 10−9, 4.3 × 10−10, and 4.4 × 10−2; the corresponding U values are provided in table S1. All KDE curves are obtained from the original distributions in fig. S10. (C) Spearman correlation coefficients between particle size and NND calculated within successive size (circles) and NND (rectangles) intervals. Arrows highlight the inferred proximity threshold and three statistically enriched size intervals for PMC sintering. (D) Difference Δsize between the size of a central particle and the mean size of its neighboring particles within 10 nm for the three identified PMC intervals (n = 191, 111, and 75 independent central particles for PMC-1, PMC-2, and PMC-3, respectively). (E) SD σsize of neighboring-particle sizes within 10 nm for the same central particles. For both (D) and (E), boxes show the 25th to 75th percentiles, whiskers span the 10th to 90th percentiles, center lines indicate the median, and dots indicate the mean. (F) Schematic illustrations of two inferred PMC-prone configurations. First, beginning at 300°C, the NND distribution develops a pronounced dip (denoted α) centered at ∼6.2 nm, flanked by two peaks β and γ. Given that PMC events also emerge after 300°C in previous ex situ HAADF-STEM observations (Fig. 2B), this may suggest that the formation of α is associated with PMC. We therefore hypothesize that the center of α represents the critical proximity threshold. Below this threshold, particles likely exhibit higher mobility and an increased probability of coalescence. This interpretation is supported by the NND distribution at 200°C (Fig. 3A), where the particles are relatively stable and less likely to sinter. The distribution is near normal, lacking sharp clustering features (dips or peaks), which indicates a largely random spatial arrangement before sintering. Therefore, this transition from a near-normal to a nonuniform distribution likely reflects particle migration induced by heating. Under this hypothesis, the enrichment at β and γ can be rationalized as follows. Upon heating and PMC are thermally allowed, particles with NND lower than α migrate toward their nearest neighbors before coalescence, reducing their NND and shifting them toward β. Enrichment at γ may also arise from PMC dynamics through two possible scenarios. In the first scenario, configurations initially around β complete PMC, and the resulting loss of neighbors increases their NND beyond α. Alternatively, NND may also increase when a particle’s nearest neighbor is attracted toward other particles and moves away. Given the asynchronicity observed in the previous section, it is reasonable that the different kinetic states represented by α, β, and γ coexist simultaneously. Conversely, clustering in the NND distribution also supports the presence of pseudotime in ex situ observations, validating PEM’s feasibility. Collectively, such β-α-γ structure suggests that α corresponds to a kinetic depletion region. Particles in α are thermodynamically unstable and thus have a short residence time before evolving toward either β or γ. Consequently, α is continuously depleted, while β and γ are correspondingly replenished by particles transitioning from α. When projected onto ensemble [or collective (22)] statistics of the representing properties (e.g., NND here), this kinetic depletion appears as a reduced probability of observation. Here, this reduction manifests as the density depletion (i.e., α) and enrichment (i.e., β and γ) in the KDE distribution. The second notable observation is that α increases with temperature. As shown in Fig. 3A, α progressively shifts toward larger NND values with increasing temperature. Its center increases from ∼6.2 nm at the onset to ∼6.3 nm at 400°C, ∼7.2 nm at 500°C, and ∼8.0 nm at 600°C. This monotonic increase reflects the temperature dependence of particle mobility. At higher temperatures, enhanced particle migration requires larger interparticle distances to prevent PMC sintering. Therefore, this monotonic increase is consistent with thermodynamic expectations and further supports α center as a critical proximity threshold for PMC. To further test this critical proximity hypothesis, we next examine particle size as an additional property dimension. If β enriches configurations prone to PMC but not yet coalesced, then particles in this population should remain relatively small. In contrast, γ contains sintered particles that have already coalesced and should therefore be larger than those in β. To test this inference, the α thresholds in Fig. 3A were used to partition particles into two groups based on whether NND was below or above the threshold. Figure 3B and fig. S10 show the particle-size KDEs for the two groups, both shifting toward larger values with increasing temperature, consistent with sintering. However, the centers remain distinct, and the differences are confirmed statistically significant (P < 0.05) by the Mann-Whitney U test across temperatures (Fig. 3B and table S1). This separation further supports α as a critical proximity threshold. If the threshold were unrelated to PMC, the resulting subpopulations would not differ significantly. Mean values further show that particles with NND > α are larger than those with NND < α at all temperatures. This pattern is consistent with the interpretation of β and γ: Peak β consists mainly of small particles that have not yet sintered, whereas γ corresponds to larger particles that have already coalesced. The size distinction between the two groups becomes more pronounced at higher temperatures. For example, the 600°C distribution in Fig. 3B broadens and shifts toward larger values relative to lower temperatures. This thermally amplified behavior mirrors the temperature-dependent shift of α in Fig. 3A. Together, the consistency between particle size and proximity further supports the existence of a critical proximity. To further examine this coupling between size and proximity, we performed a direct correlation analysis. Using the 400°C window (closest to TTammann), NND values were grouped into 1-nm bins, and the Spearman correlation coefficient between size and NND (ρsize,NND) was calculated for each bin. As shown in Fig. 3C, the two properties are positively correlated when NND is below ∼6 nm, consistent with sintering progression simultaneously altering both size (via OR) and proximity (via PMC). In contrast, once NND exceeds ∼6 nm, ρsize,NND fluctuates around zero, indicating a weak correlation. This regime likely corresponds to particles that either have not undergone PMC or have already coalesced, where size changes become largely independent of proximity. Notably, this turning point (∼6 nm) coincides with the threshold identified from Fig. 3A (∼6.3 nm). Beyond proximity, the correlation analysis also suggests potential size preferences for sintering. Figure 3C shows that ρsize,NND exhibits three local dips corresponding to three size intervals: small (PMC-1, 2 to 3 nm2), medium (PMC-2, 6 to 7 nm2), and large particles (PMC-3, 8 to 9 nm2). By analogy with the turning point observed across NND bins, these inflections may indicate unusual size-proximity coupling, suggesting that particles in these intervals are more likely to participate in restructuring events leading to sintering. To further characterize these preferences, the difference between each particle’s size and the average size of its neighbors within 10 nm (Δsize) was quantified. As shown in Fig. 3D, Δsize is predominantly negative for PMC-1 (mean, −3.65 nm2), indicating that small particles are typically surrounded by larger neighbors. Correspondingly, Δsize is mainly positive for PMC-3 (mean, 1.53 nm2), consistent with this large-small configurational preference. By contrast, PMC-2 shows Δsize values centered near zero, with a mean of 0.07 nm2, indicating interactions with both larger and smaller neighbors as well as clustering among similarly sized particles. This interpretation is further supported by the SD of neighboring particle sizes (σsize) shown in Fig. 3E. PMC-2 exhibits the highest uniformity among the three intervals, with σsize = 0.99 nm2, lower than PMC-1 (1.32 nm2) and PMC-3 (1.88 nm2), suggesting a higher probability of size homogeneous configurations. Together, these observations suggest two preferred PMC configurations in Fig. 3F: (i) coalescence between large and small particles in Fig. 3Fa and (ii) coalescence among particles of similar medium sizes in Fig. 3Fb. Nevertheless, particle size effects should be interpreted more cautiously than proximity. Size primarily influences OR rates through its effects on surface chemical potential and atomic mobility, rather than acting as a universal trigger for bulk mobility. Proximity effects, by contrast, directly control interparticle attraction and thus coalescence probability. Together, these results suggest a plausible scenario: Once proximity falls below the threshold, particles within preferred size intervals evolve through OR, approach each other, and eventually coalesce. In summary, we conclude that the rationality of PEM arises from the transient nature of unstable restructuring states (e.g., particles with an NND around the critical threshold). When observed along a pseudotemporal axis defined by an appropriate structural property, such transient states appear as probability-density depletion in the collective distribution of this property, while adjacent states become correspondingly enriched. In this sense, these observable depletion and enrichment patterns are the statistical signatures of different stages of the sintering process, thereby underpinning the feasibility of PEM. It is precisely by leveraging this principle that PEM reveals proximity as the trigger-like NND controlling PMC sintering in Au/SiNx and quantifies its thresholds. In situ sintering validation To validate the above sintering inferences by PEM, we performed in situ HAADF-STEM imaging on a fixed ROI under identical conditions. Figure 4A shows the ROI at 200°C, below THüttig, at which both severe OR and PMC are not expected, as we mentioned before. We applied two key priors derived from PEM at 400°C (where severe PMC are thermodynamically allowed to occur): (i) Particles with NND < 6.3 nm are highly susceptible to PMC; and (ii) particles within preferred size intervals (e.g., PMC-1, PMC-2, and PMC-3) may have elevated PMC probability. Thus, Au particles satisfying NND < 6.3 nm were color-mapped (with the help of NanoSeg) in Fig. 4A according to their size intervals. Among these early warnings, three of the four particles labeled (a), (b), (c), and (d) in the configuration B shown in Fig. 4A were predicted to undergo PMC, and it was therefore selected for focused tracking. In situ validation of pseudotime inferences. (A) HAADF-STEM image of the in situ tracking region of Au/SiNx at the initial 200°C. (B to G) Sequence of in situ HAADF-STEM images tracking the sintering process of a four-particle configuration during heating from 200° to 400°C. Particles satisfying NND < 6.3 nm (inferred in Fig. 3A) are color-mapped according to three PMC intervals in Fig. 3C. (B) is an enlarged view cropped from (A). The scale bar in (B) applies to (B) to (G). (H and I) Configurations H and I, initially inferred to be stable, are predicted to coalesce during heating. (J) Configuration J with NNDs exceeding 6.3 nm, despite having PMC-prone particle sizes, remains stable after 400°C for 30 min. (K) Time evolution of the particle population within the threshold-centered depletion regions during in situ heating, overlaid with the temperature profile. (L) Thresholding curve of sintering prediction accuracy as a function of the critical NND threshold at the 300° to 400°C interval. The enlarged view highlights the prediction accuracy at the 6.3-nm threshold, reaching 92% and approaching the theoretical predictive upper bound of 94% obtained by exhaustive threshold sweeping. The inset shows the confusion matrix of 6.3-nm–based classification. In Fig. 4B, particle (a) (2.4 nm2) and particle (b) (9.4 nm2) belong to PMC-1 and PMC-3, respectively. The (a) to (b) distance (3.1 nm) is shorter than the (b) to (c) distance (4.7 nm), suggesting that (a) and (b) are most likely to coalesce first. Upon heating to 300°C (Fig. 4C), OR sintering had occurred, particle (a) shrank from 2.4 to 2.0 nm2, while (b) grew from 9.4 to 10.1 nm2. Meanwhile, (c) moved closer to (b) from 4.7 to 4.5 nm, suggesting that (c) may coalesce subsequently. After 61 s in Fig. 4D, (a) and (b) coalesced as predicted, forming a 10.4-nm2 particle, supporting the mechanism in Fig. 3Fa. Notably, this event occurred at 300°C, well below the bulk TTammann of 396°C, which can be rationalized by the fact that TTammann is substantially reduced at the nanoscale (6). This suggests that, even at temperatures lower than the theoretical threshold, small particles such as (a) may become mobile and preferentially coalesce with immobile but larger neighbors. As temperature increases, the size threshold for particle mobility loosens, implying that medium-sized particles (e.g., PMC-2) may subsequently be allowed to participate in PMC. Right after the first coalescence, priors from PEM also warned of further coalescence among (b), (c), and (d) in Fig. 4D. We therefore propose that relatively large particle (d) would need to shrink to become mobile [likely via OR with (c)], resulting in size homogenization and reduced distance, while particle (c) would also approach the immobile (b). These inferences were confirmed later at 400°C in Fig. 4E. Particle (d) decreased from 9.0 to 7.9 nm2, which is abnormal in sintering, but approached (c) from 6.0 to 3.3 nm. Meanwhile, particle (b) grew to 10.8 nm2, and (c) decreased to 4.9 nm2 and approached (b) to 3.7 nm. It is likely that, among them, (c) underwent the coupled ripening of both (d) to (c) and (c) to (b). Nonetheless, the size of (c) increased, indicating that the substantial mass transfer from (d) exceeded the loss to (b). This phenomenon may suggest a potential inverse growth mechanism, where a particle may first undergo shrinkage through ripening to achieve mobility before coalescence. This mechanism may complement the conventional monotonic growth picture. Moreover, this (c) and (d) size homogenization observed from Fig. 4 (D and E) (SD among them decreased by 1.5 nm2) supports the mechanism in Fig. 3Fb. Continued heating at 400°C, Fig. 4F records the moment at 41 s when (c) merged into (b) (due to top-to-bottom raster scanning in HAADF-STEM). At 91 s, particle (d) also coalesced. From Fig. 4 (A to G), the full fate of particles (a) to (d) unfolded precisely as predicted by PEM. It is found that PEM could provide real-time warnings during the dynamics rather than merely initial ones, as particle (d) began to be flagged in Fig. 4D. This dynamic predictive capability was further validated on configurations H and I in Fig. 4A. Initially not expected to coalesce at 200°C, they were subsequently flagged at 300°C (Fig. 4H) and 400°C (Fig. 4I), respectively, and ultimately coalesced. Moreover, in configuration H, particle sizes were homogenized from 6.9 ± 1.3 nm2 (Fig. 4A) to 6.3 ± 0.6 nm2 (Fig. 4H, within PMC-2) before coalescence (to 12.8 nm2). Two similarly sized particles in configuration I (7.6 and 7.1 nm2, near PMC-2) also coalesced preferentially (to 12.2 nm2) despite the presence of a larger neighbor (10.3 nm2). Together with the (c) and (d) homogenization observed in configurations B (Fig. 4, D and E), the results of configurations H and I reinforce the proposed mechanism in Fig. 3Fb. However, it should be emphasized that realistic sintering is intrinsically governed by the coupled evolution of interparticle proximity and intraparticle morphology. Throughout the above fate tracking, they almost always changed simultaneously. Nevertheless, their roles may be distinct. Configuration J in Fig. 4A provides a critical illustration: Although its three particles initially are PMC prone sized (8.2, 8.4, and 8.8 nm2, all within PMC-3 interval), their NND (9.8 and 11.2 nm) substantially exceeded the 6.3-nm threshold suggested by PEM. After 400°C for 30 min (Fig. 4J), no PMC occurred. Instead, separations increased to 10.7 and 12.4 nm. This observation suggests that proximity may function as a necessary condition for triggering a PMC process, while specific size may further accelerate this process. Given the trigger-like role of proximity revealed above, it can be leveraged as a criterion for predicting whether an individual particle will undergo PMC sintering. To validate this predictive capability, we first examined the approximate kinetic depletion region around the critical threshold (α ± 0.5 nm) and tracked the population (or count) of Au particles within this interval during in situ heating. The time evolution of this particle population is shown in Fig. 4K, overall demonstrating fluctuations across all temperature windows followed by a marked decrease upon the onset of severe sintering (∼340°C). Specifically, during the 300° to 400°C session, where the temperature gradually approached and exceeded the TTammann of Au (396°C), the total particle number decreased by 40%, far exceeding the reductions observed from 200° to 300°C (−12%) and 400° to 500°C (−5%), consistent with our previous expectation that sintering is most severe in this session. Correspondingly, the particle population within the depletion region dropped markedly from an average of 3.7 (200° to 300°C, 6.2 ± 0.5 nm) to 1.5 (300° to 400°C, 6.3 ± 0.5 nm) across observation time points, with only 1.2 after reaching 400°C (>TTammann). This low population (count) indicates that, once large-scale sintering begins, particles within this NND range are rapidly consumed, reflecting their kinetic instability and confirming our previous hypothesis of the depletion-enrichment mode. Specifically, these particles either approach and coalesce or move apart, resulting in the α depletion and β/γ enrichments in Fig. 3A. Minor fluctuations could be attributed to transient replenishment (or dynamic balance) as particles intermittently entered the depletion region but were rapidly consumed, resulting in a persistent low population that can even occasionally approach zero (Fig. 4K), indicating that depletion outpaces replenishment. Moreover, this low population persisted into the subsequent higher-temperature 400° to 500°C session (7.2 ± 0.5 nm), also consistent with the previous inference that the critical threshold would shift toward larger NND with increasing temperature. This validated depletion phenomenon provides kinetic rationality to directly support the predictive capability of the critical proximity in sintering. To quantitatively evaluate whether this proximity effect can be leveraged as a predictive handle to mitigate PMC sintering, we further focus on the 300° to 400°C session, where PMC is most pronounced (−40%) as mentioned. Using the PEM-derived threshold of 6.3 nm as the criterion, all 63 Au particles observed at the initial state (300°C, 0 min) were divided into two groups (figs. S11 to S19): (i) 27 particles with NND < 6.3 nm were predicted to sinter (i.e., positive), and (ii) 36 particles with NND > 6.3 nm were predicted not to sinter (i.e., negative). These particles were then continuously tracked (figs. S11 to S18) until the end of the session (400°C, 39 min). As shown in fig. S19, of the 27 positive predictions, 26 were true positives, confirmed by observed PMC events with neighboring particles during heating, while only one was a false positive (no sintering occurred). Of the 36 negative predictions, 4 were false negatives, whereas the remaining 32 were true negatives and did not undergo PMC sintering. Accordingly, the proximity-based prediction achieves a precision of ∼96% (fraction of predicted-sintering particles that do sinter), a recall of ∼87% (fraction of all sintering particles correctly identified), and an overall accuracy of ∼92% (fraction of all particles correctly classified). Notably, this accuracy already approaches the theoretical predictive upper bound of NND itself (∼94%), obtained by exhaustive threshold sweeping against the in situ observed ground truth (Fig. 4L). Moreover, the PEM-inferred 6.3 nm differs from the optimal threshold of 6.1 nm by only 2 Å (∼+ 3%). In summary, this direct quantitative assessment establishes the critical proximity as a predictive criterion for PMC and supports its use as a design parameter for suppressing catalyst sintering. It confirms the capability and accuracy of PEM to identify kinetic control factors through statistical ordering of restructuring states. Collectively, the sintering inferences from PEM in the previous section are directly supported by in situ HAADF-STEM tracking. Therefore, we conclude that, with the leverage of pseudotime, even static snapshots of ex situ electron microscopy could suggest how and when sintering occurs, rather than simply comparing structure changes between the two time points of before and after reaction (i.e., fresh and aged). Such fundamental knowledge of the restructuring processes can be translated into the rational development of more stable catalysts that can better withstand such processes, thus improving stability. Pseudotime transferability With the validity of PEM established, its transferability to practical catalyst systems is next considered. The preceding results indicate that pseudotime reconstruction rests on three requirements: structural heterogeneity [or nonuniformity (5)], appropriate sampling timing, and statistical representativeness. First, structural heterogeneity is needed to generate asynchronous evolution but is unlikely to be a limiting condition in supported catalysts, where perfect uniformity in all aspects is rarely expected (2, 5, 42). Second, appropriate timing requires restructuring to have already occurred, at least in part, before ex situ observation. In practice, this can be ensured by controlling the dominant driving force, for example, the temperature (e.g., TTammann here as a reference) in thermal sintering, or analogously, the overpotential in electrocatalysis and light intensity in photocatalysis. In contrast, the third factor, sampling bias (5, 36), emerges as the primary limitation to transferability. Subsampling (independent and identically distributed) ex situ HAADF-STEM images and repeating the pseudotime analysis yield substantial deviations from the full-dataset result (48 distributed ROIs), as shown in figs. S20 to S23. For the Au/SiNx system here, at least ∼300 particles, or ∼10 ROIs (snapshots), are required to maintain robustness. This requirement is expected to increase substantially for practical systems, where morphology, contrast, and noise are more complex or heterogeneous. Although batched ex situ imaging across multiple regions is experimentally feasible, the bottleneck shifts to high-throughput micrograph processing. Reliable PEM requires accurate extraction of structural descriptors (e.g., proximity and size in this study) at the single-particle level, yet manual measurements (e.g., via ImageJ) and annotation-intensive models (e.g., conventional U-Net) remain tedious and error-prone. Transferability of PEM therefore calls for an automated and annotation-efficient structural quantification. Accordingly, we developed NanoSeg as an enabling implementation for automated micrograph segmentation and subsequent extraction of the particle-level descriptors required for pseudotime reconstruction. Its segmentation of Au/SiNx is illustrated in Fig. 5A, while the accuracy of the resulting structural descriptors (e.g., size and NND) is quantitatively validated in fig. S24 and table S2; methodological and validation details of NanoSeg are provided in Supplementary Text. Transferable construction of pseudotime. Representative HAADF-STEM images of Au/SiNx (A) and PtSn alloy nanoparticles supported on Al2O3 (B), together with the corresponding NanoSeg segmentation results, manually annotated ground truth, and prediction differences. TP, FP, and FN denote true positives (correctly identified particles), false positives (incorrectly identified particles), and false negatives (missed particles), respectively. (B) is the same as micrograph no. 23 in figs. S25 and S38. (A) and (B) are the same as cropped patches in Fig. 1. However, practical catalysts generally exhibit greater structural heterogeneity than the model Au/SiNx system, potentially compromising segmentation accuracy. Moreover, this heterogeneity can be further compounded by imaging-induced variations, producing pronounced differences in noise, contrast, focus, and apparent morphology among catalyst entities within the same micrograph. Together, these variations can undermine reliable extraction of structural descriptors and thus pseudotime reconstruction of practical systems. Therefore, we evaluated NanoSeg on a highly challenging dataset, the HAADF-STEM images of PtSn nanoalloys supported on Al2O3 (fig. S25). This system exhibits substantial variations including size (fig. S26), shape (fig. S27), contrast (fig. S28), and focus quality (fig. S29). Despite these overlapping complexities, NanoSeg achieves an overall Dice score of 0.78 for PtSn instances (Fig. 5B, figs. S30 to S39, and tables S3 and S4), lower than that for Au/SiNx and supporting our preceding argument that greater heterogeneity can reduce the robustness of automated quantification. Limiting this impact requires not only larger sample sizes but also near-optimal segmentation accuracy, because additional sampling cannot correct subsequent biased quantification and may instead propagate that bias into pseudotime reconstruction. Accordingly, we benchmarked NanoSeg against representative approaches (figs. S40 to S61 and tables S5 to S12) and evaluated its applicability across applied catalyst systems—from single-atom catalysts to overlapping fuel-cell nanoparticles—and across imaging modes (figs. S62 to S68 and tables S13 to S18). The results demonstrate comparable performance with minimal annotation while identifying microscope artifacts and reference selection (figs. S69 to S75 and tables S19 to S26) as key considerations for practical deployment. In summary, structural heterogeneity is both the prerequisite and the bottleneck of pseudotime reconstruction. On the one hand, it induces the asynchronous evolution essential for pseudotime to emerge. On the other hand, it challenges the accurate quantification of structural descriptors, whose values encode individual kinetic states and collectively constitute the pseudotime axis. Therefore, we conclude that transferring pseudotime to practical catalyst systems requires not only appropriate timing and sufficient sampling but also accurate and high-throughput quantification of structures. NanoSeg exemplifies one such approach by minimizing additional bias during micrograph processing, thereby preserving fidelity of pseudotime. DISCUSSION The present results establish pseudotime in ex situ and show that it can reconstruct catalyst evolution. This reconstruction enables direct quantification of the critical proximity that triggers Au nanoparticle migration and coalescence. In situ HAADF-STEM confirms that particle mobility, and thus sintering, is markedly suppressed beyond this threshold. The resolved proximity effect arises from the property-restructuring states mapping established by pseudotime, while such dynamic interpretability of pseudotime originates from asynchronous aging induced by structural heterogeneity. A deep learning framework further enables automated construction of pseudotime across catalyst systems, suggesting scalability. Given that pseudotime has not yet been well explored in material characterization, we anticipate that it could supplement the general view that postreaction observation is inherently static and open a new route to probe restructuring dynamics. Despite the identified proximity effect helping design sinter-resistant catalysts, several limitations should be noted. First, the flat SiNx window support allows proximity to be approximated by planar projection, whereas realistic catalysts on topographically complex surfaces of oxides (e.g., Al2O3, TiO2, and CeO2) or within porous zeolites (e.g., ZSM-5, SBA-15, and S-1) require three-dimensional distances, obtainable via contrast-based estimation (43) or three-dimensional electron tomography (22). Second, the NND here serves as a representative of proximity. More comprehensive descriptors [e.g., higher-order neighbors (22) or even continuous pair distribution functions] remain to be explored. Third, this work focuses on PMC-dominated sintering, whereas supported metal catalysts may also undergo OR sintering, particle reshaping and refaceting, surface segregation or dealloying, phase transitions, and support-derived encapsulation under reaction conditions (2, 42). Extending pseudotime beyond sintering requires measurable features that reflect the progression of each process and help distinguish it from competing mechanisms. For OR, particle size and curvature should be considered together with the growth or shrinkage of neighboring particles, because particle growth alone cannot distinguish ripening from coalescence (7). Particle reshaping may be tracked through changes in facet exposure and overall particle shape in lattice-resolved transmission electron microscopy images (44), whereas segregation and dealloying require spatially resolved composition from STEM-EDS or STEM–electron energy-loss spectroscopy (EELS) spectrum imaging (45). Phase transitions may be followed through phase-specific lattice patterns or electron diffraction peaks (46), while support-derived encapsulation may be tracked through overlayer coverage and thickness, with EELS mapping confirming the composition (47). Together, these considerations suggest that extending pseudotime beyond sintering requires selecting features that are most characteristic of, and evolve directly with, the specific restructuring process, rather than relying on proximity here alone. As a proof of concept, pseudotime should also be interpreted with several general caveats. First, it remains a retrospective inference from postreaction observations and may miss certain transient events, such as reported back-relaxations (48, 49) upon removing reaction conditions. This limitation may be alleviated by combining ex situ snapshots with a few in situ measurements to allow mutual calibration or by linking ex situ observations across changing conditions (e.g., atmosphere, pressure, and temperature), to reduce bias from single-point observation. Second, pseudotime relies on intrinsic structural heterogeneity to induce asynchronous aging. If the catalyst system is highly uniform or evolves synchronously, then the recoverable dynamic information may be reduced, weakening reconstruction. Although such strict homogeneity, as we mentioned earlier, is uncommon for metal catalysts under operating conditions (2, 5, 42), defining applicability criteria remains valuable. For instance, entropy-based metrics of heterogeneity may provide a way to quantify identifiability (or applicability) of pseudotime. Overall, these limitations do not affect the core validity of pseudotime but define the scope for its future development. As an underexplored analytical concept (or scale) in material characterization, pseudotime, either independently or in concert with existing known techniques, may enable more efficient and refined analysis of catalyst structural evolution. MATERIALS AND METHODS Micrograph acquisition All electron microscopy image data and manual annotations (if available) used in this work were accessed from publicly available datasets. HAADF-STEM images of Au/SiNx, PtSn/Al2O3, Pt/C, and BF-TEM images of Au/ZSM-5 were obtained from (50) under the accession code: https://doi.org/10.5281/zenodo.14791390. BF-TEM images of Pd/C were obtained from (51) under the accession code: https://doi.org/10.5281/zenodo.7024892. ABF-STEM images of PtCo/C were obtained from (52) under the accession code: https://github.com/SKKU-STEM/PMA-Net. HAADF-STEM images of Pt/NC were obtained from (53) under the accession code: https://github.com/HPAI-BSC/AtomDetection_ACSTEM. Micrograph segmentation NanoSeg segmentation and benchmarking were implemented in PyTorch and accelerated through CUDA on an L40S GPU. Within NanoSeg, a single annotated reference image was first used to train a catalyst feature extractor. Feature correlation across test images was used to generate saliency maps. These maps were further enhanced through contrast normalization and filtered to produce both bounding box and point prompts for individual catalyst instances. The generated prompts, together with the corresponding images, were then fed into SAM (54) to obtain binary masks. Nonmaximum suppression was applied to remove overlapping detections. NanoSeg details are provided in the Supplementary Materials.