Chemicals and solvents involved in this work were purchased and used without further purification. Platinum acetylacetonate (C₁₀H₁₄O₄Pt) was purchased from Bide Pharmatech. Zinc acetylacetonate (C₁₀H₁₄ZnO₄), sodium tetrachloropalladate (Na₂PdCl₄), chloroauric acid (HAuCl₄) and 2-methylimidazole (C₄H₆N₂) were purchased from Aladdin. Graphene oxide (GO) was purchased from Tanfeng Technology. Polyvinyl pyrrolidone (PVP) was purchased from Macklin. Zinc nitrate hexahydrate (Zn(NO₃)₂·6H₂O), sulfuric acid (H₂SO₄), nitric acid (HNO₃), acetone, methanol and ethanol were purchased from the Sinopharm Chemical Reagent Co. Ltd.

GO was pretreated in a mixed solution of H₂SO₄ (10 mL) and HNO₃. C₁₀H₁₄O₄Pt (22.4 mg), C₁₀H₁₄ZnO₄ (11.2 mg), and GO (48 mg) were dispersed in a mixture of ethanol (40 mL) and acetone (40 mL). The mixture was sonicated for 1 h and subsequently stirred at 300 rpm to dryness in a water bath at 30 °C for approximately 6 h. The resulting dry powder was annealed in a tube furnace at 700 °C for 2 h under 10% H₂/Ar flow. The final product was designated as PtZn/GO.

MNPs were synthesized by the reported synthesis method^[18,37]. MNPs were capped by PVP (Mw = 55,000) first as reported methods. Then 1 mL MNPs solution, 5 mL solution of 2-methylimidazole (25 mM, in methanol), and 5 mL solution of Zn(NO₃)₂•6H₂O (25 mM, in methanol) were mixed. The solutions were then reacted for 24 h without stirring at room temperature. The product was collected by centrifugation, washed several times with methanol.

The morphology, structure and elemental information of materials were observed on transmission electron microscope (TEM, JEOL JEM-F200 200kV, Japan) coupled with EDS (JEOL Super Dual EDS System, Japan). The XRD was recorded using an X-ray diffractometer (Rigaku Miniflex600, Japan) with a Cu-Kα radiation source (λ = 1.54 Å). Ultrasonication was performed using an ultrasonic cleaner (KQ-400DE, Kunshan Ultrasonic Instruments Co., Ltd., China). Stirring was performed using a magnetic stirrer (MS-H380-Pro, DLAB Scientific Co., Ltd., China). Centrifugation was carried out using a centrifuge (TG16-WS, Hunan Xiangyi Laboratory Instrument Development Co., Ltd., China). The annealing process was performed in a tube furnace (ZHK-G05123K, Tianjin Zhonghuan Electric Furnace Co., Ltd., China).

We performed a Gaussian filter on several TEM images to remove the image noise. Subsequently, the images were binarized using the Otsu thresholding algorithm to select the region of MNPs. Connected regions were segmented using the watershed algorithm, with only regions exhibiting a circularity >0.7 selected for statistical analysis. The mean value of the major and minor axes of bounding box for each isolated region was calculated as the particle diameter. We employed this procedure to calculate the particle size distribution of Au NPs (15.70 ± 1.80 nm) and Pd NPs (15.40 ± 2.12 nm), as shown in Supplementary Figure 1.

The same pre-processing procedure were utilized on all tilt series of sample, including intensity normalization, image alignment, background subtraction. Subsequently, 200 cycles of simultaneous iterative reconstruction technique (SIRT) algorithm were performed to reconstruct the 3D volumes, followed by volume segmentation to extract structural information. To minimize the impact of missing wedges and other reconstruction artifacts on particle localization, we employed a three-dimensional Gaussian peak fitting method for particle localization^[38]. Local maxima were first determined for all particles in the 3D volume data, and then the particle center positions were determined by the polynomial fitting method. In addition, we removed weak intensity positions and repetitive positions by setting limits on the integral intensity of the 3D box centered on the local maxima and on the minimum distance between the two local maxima, respectively.

We used the measured distance between two particles (d), the average of particle size (μ) and mean square of distribution (σ) to determine whether MNPs were agglomerated [Supplementary Figure 2A]. From a statistical perspective, for NP particle sizes that follow a near-normal distribution, the range of μ ± 2σ encompasses approximately 95% of the particles. If d ≥ μ + 2σ, two particles are recognized as non-aggregated. If μ - 2σ < d < μ + 2σ, two particles are recognized as aggregated. μ-2σ is taken as the minimum distance for particle tracking. Since a cutoff value for the distance has been selected, the minimum resolution distance for this analysis is μ-2σ.

To determine whether the MNPs are encapsulated in the ZIF-8 shell, we first identified the ZIF-8 shell with the intensity threshold and further extracted its surface layer position [Supplementary Figure 2B]. The distance of each particle relative to the shell layer was calculated as d_depth. If $$ d_{\text {depth }}<\frac{\mu+2 \sigma}{2} $$, the particles are considered to be encapsulated. Else if $$ d_{\text {depth }} \geq \frac{\mu+2 \sigma}{2} $$, the particles are considered to be unencapsulated.

The three-dimensional centers of all NPs were localized, and the distance from each NP to the center was calculated. Particles within each concentric shell were counted and normalized by the corresponding shell volume. The shell thickness was set to 25 nm, and the innermost shell began at a radial distance of 50 nm from the center (to avoid large counting errors arising from the small volume near the center).

We propose the following items to systematically accelerate ET data acquisition: First, within or near the region of interest (ROI), the image shifts in the x- and y- directions along with defocus values are rapidly recorded at coarse tilt increments. These datasets are then used to fit the function of stage shifts and defocusing. Subsequently, the actual tilt series is acquired at fine angular increments with automated correction [Figure 1A and B]. Using this method, the total ROI acquisition time is reduced from 1 h 20 min to 9 min 54 s with high automation. Considering an electron beam in annular dark-field scanning TEM mode with a 50 pA beam current^[39], the estimated total electron dose is reduced from 8.90 × 10³ e/Ų to 988.7 e/Ų - representing an 88.9% reduction, see details in Supplementary Table 1. The image difference of the projections at 0° before and after conventional/fast methods demonstrates that the fast method causes significantly less damage to MNPs@ZIF-8 samples than the conventional method [Figure 1C]. When MNPs@ZIF-8 is subjected to prolonged electron beam irradiation, the sample is at risk of damage and rotation, which constitutes the primary cause of ET reconstruction failure [Figure 1D(i)]. In contrast, the 3D volume reconstructed from fast ET acquisition method shows higher spatial accuracy [Figure 1D(ii)].

We specifically discussed the advantages of fast ET acquisition in reducing image drift and defocusing. Compared to the no-calibration method, the pre-calibration acquisition significantly minimized both image shifts (the maximum from ~ 2 μm to < 150 nm) and defocus values (the maximum from ~ 3 μm to < 250 nm), as shown in Figure 2A and B. To further investigate the impact of different calibration positions, we manually recorded multiple tilt-series datasets include xy-shifts and defocus values from reference position (A₀), the identical position (A_1~3), and two near positions (B for 2 μm offset, C for 10 μm offset), with 4° increments across ±65°, 33 in total, as shown in Figure 2C. The test results of stage xy-shifts offsets and defocus demonstrate reproducible behavior at both identical and near positions [Figure 2D-F]. Without pre-calibration, image shifts in x-direction (parallel to the tilt axis, the shift maximum is about 300 nm) are consistently smaller than those in y-direction (perpendicular to the tilt axis, shift maximum ~ 2 μm), consistent with previous studies^[40].

In order to quantify the reproducibility of stage shifts and defocus values, we calculated x-shift errors, y-shift errors and defocus errors of A_1~3, B and C, presenting the results as box plots [Figure 2G-I, see details in Supplementary Tables 2-4]. The results demonstrate that stage shifts and defocus values acquired from near positions (B and C) exhibit larger errors compared to those collected at identical positions (A_1~3). These errors increase progressively with the distance between the reference (A₀) and test positions. The box plots also indicate that, with few exceptions, xy-shift errors in identical position acquisitions and near position B are constrained to within 200 nm. These errors primarily originate from mechanical vibrations of stage during tilting, which fundamentally limit the minimum acquirable window size in pre-calibration methods. Without additional manual adjustment, a margin of at least 200 nm must be reserved on all sides of the acquisition window to prevent sample drift beyond the imaging area. It is worth mentioning that data from position B exhibits consistent positive y-direction shifts relative to A₀ between 25°-60° [Figure 2E]. A minor manual adjustment (applying a -80 nm image shift in the y-direction at 25°) can significantly reduce y-shift errors from [-55.0 nm, 128.9 nm] to [-55.0 nm, 54.1 nm], as shown in Supplementary Figure 3. This suggests that simple manual adjustments could further decrease the necessary margins of this pre-calibration method. It is worth noting that the safety window may vary depending on the mechanical stability of different electron microscope sample stages.

After verifying the reproducibility of stage shifts and defocus during tilting, we evaluated the types and parameters of the fitting functions. Pre-calibration data, consisting of a subset of the A₀ or B data, are used to fit multiple polynomial and interpolation functions that predict image shifts and defocus values at other tilt angles. The root mean square error (RMSE) between A₃ data and predicted data is utilized to assess the prediction accuracy of the fitted functions[Figure 2J-L]. The subsets are constructed using anchor points selected at varying angular increments (8°, 12°, 16°, 20°, 24°). The results reveal that: (i) For both identical-position and near-position predictions, the RMSE of x-direction functions falls within 25-60 nm; (ii) y-direction prediction functions exhibit RMSE of 5-40 nm (identical-position) and 60-90 nm (near-position) respectively; (iii) Defocus value predictions show RMSE ranging from 40-80 nm (identical-position) and 90-115 nm (near-position). Considering both inherent prediction inaccuracies and reproducibility errors across tilt series, these observed deviations fall within expected ranges. Fitting functions utilizing more anchor points (or smaller angular increments) demonstrate improved prediction accuracy, particularly in the y-direction. Regarding the type of fitting function, interpolation methods generally outperform with abundant anchor points, while polynomial fitting exhibits better stability with fewer anchor points. In addition, the images acquired using the rapid method exhibit sufficiently high integrity and clarity, which validates the effectiveness of sample motion and defocus (see Comparison of fast and conventional acquisition Section, and Supplementary Figures 4 and 5 in Supplementary Materials for details).

The standard Au NPs@ZIF-8 is synthesized following established methods^[17], and characterized by TEM, energy dispersive X-ray spectroscopy (EDS) mapping, and powder X-ray diffraction (XRD), as shown in Supplementary Figure 6. Subsequent synthesis is performed following a controlled-variable approach, yielding various samples as summarized in Supplementary Table 5. The developed fast ET method is employed to study the 3D architecture of MNPs@ZIF-8 (see MNPs@ZIF-8 series acquisition Section in Supplementary Materials for details), thereby establishing a direct relationship between synthesis parameters and 3D structural characteristic. Comparison of 0° projections before and after fast ET acquisition shows no detectable structural damage, confirming ET acquisition reliability [Supplementary Figure 7]. The dispersed MNPs (higher voxel intensity) are embedded in a ZIF-8 matrix (lower voxel intensity), and are observed both incorporated inside the matrix and attached to the surface, with distinct aggregations present in specific regions [Figure 3A]. The 3D coordinates of MNPs are determined from the 3D volumes based on the voxel intensity, enabling quantitative assessment of MNPs aggregation ratio and encapsulation efficiency [Figure 3B]. The spatial distribution patterns of MNPs are further quantified through NND distribution fitting and RD fitting, as shown in Figure 3C. The developed 3D voxel analysis workflow is universally applicable for structural investigations of diverse samples with various synthesis conditions, as shown in Figure 4.

Quantitative analysis of reference Au NPs@ZIF-8 sample reveals complete encapsulation of all 219 Au NPs within ZIF-8 matrix, with MNPs aggregation ratio of 12.8% [Figure 4A]. In Comparison to the reference sample, the Pd NPs@ZIF-8 sample exhibits similar MNPs distribution characteristics [Figure 4B], with comparable MNPs encapsulation efficiency, dispersion quality, and analogous RD and NND patterns. These results indicate that the type of MNPs has a limited effect on the 3D architecture of sample. This observation aligns with the known mechanism where Zn²⁺ sites on fresh ZIF-8 surfaces adsorb polyvinylpyrrolidone (PVP) capping agents of MNPs@PVP^[6,10], thereby rendering MNPs dispersion largely independent of the MNPs core composition.

Extension of the synthesis time to 48 h results in the appearance of unencapsulated MNPs [Figure 4C]. Since the growth of ZIF-8 is basically completed within the standard reaction time, the ions and ligands in the solution are insufficient to support the growth of new ZIF-8 shells. Subsequent reaction leads to the adsorption of residual free MNPs onto the ZIF-8 surface. Interestingly, along with unencapsulated MNPs, several MNPs are found to be shallowly encapsulated (orange arrows in Figure 4C). These shallowly encapsulated MNPs, which are also prevalent in other samples exhibiting external MNPs, are difficult to detect by conventional 2D projection techniques. In solution environments, ripened ZIF-8 undergoes dissolution-recrystallization dynamics^[40], enabling partial internalization of surface-adsorbed MNPs into the ZIF-8 matrix to form shallow encapsulation structures.

We also investigated the influence of adjusting MNPs concentration and Zn²⁺: ligand ratio. At the moderately increased concentration (1.15×, named mMNPs1), both the quantity and encapsulation efficiency of MNPs remain comparable to the reference sample. In contrast, the higher concentration (2×, named mMNPs2) exhibits both greater abundance of encapsulated MNPs and the emergence of unencapsulated MNPs [Figure 4D and E]. Samples with higher Zn²⁺: ligand ratio exhibited similar encapsulation and agglomeration phenomena [Figure 4F]. Previous studies have indicated that an increased salt-to-ligand ratio suppresses the final yield of ZIF-8 and increases the particle size of ZIF-8^[41]. This results in an increase in the relative concentration of MNPs when the Zn²⁺: ligand ratio is increased.

We further investigated the effect of extra PVP addition: a sample with an extra 1 × 10^-4 wt.% PVP addition (denoted as mPVP1) and another with an extra 5 × 10^-4 wt.% PVP addition (denoted as mPVP2), as shown in Figure 4G and H. The extra addition of PVP significantly reduces the aggregation ratio of MNPs, and this trend became more pronounced as the PVP addition increase [Figure 4I]. Unencapsulated particles are also detected, and their fraction increases proportionally with the PVP concentration. Previous studies have indicated that the introduction of additional PVP molecules competitively occupies the adsorption active sites originally available for MNPs@PVP^[42]. This leads to a sparse distribution of adsorption sites on the ZIF-8 shell, thereby causing the MNPs to exhibit enhanced dispersibility.

Insight into the radial characteristics of MNPs is essential for the mechanistic understanding of the encapsulation process. We conduct a discussion on the RD of all the above-mentioned samples, and the results reveals two distinct distribution patterns [Figure 4J]. For Type I (Ref., Pd, Time and mMNPs1), the RD fitting curves follow an exponential decay pattern, as shown in Supplementary Figure 8A-D. This RD features correlates with the temporal decrease in MNPs concentration during synthesis. During the initial nucleation stage of ZIF-8, the high concentration of free MNPs exhibits higher adsorption probability. As ZIF-8 growth progresses, the concentration of available MNPs decreases, leading to a corresponding reduction in adsorption probability. For Type II (mMNPs2, Ratio, mPVP1 and mPVP2), The fitting curves are better described by a piecewise function combining a saturation plateau with exponential decay [Supplementary Figure 8E-H]. The detailed parameters of the RD fit are shown in Supplementary Tables 6 and 7. The observed plateau in the RD profiles suggest saturation behavior in the MNPs adsorption process. All four conditions in Type II are designed to increase the relative concentration between MNPs and adsorption sites. Specifically, mMNPs2 achieves this by raising the MNP concentration, while Ratio, mPVP1, and mPVP2 work by reducing the concentration of ZIF-8 adsorption sites. This saturation of adsorption sites during the initial stage generates a characteristic plateau in the RD profile. As the reaction progresses, the continuing growth of ZIF-8 provides an increasing number of adsorption sites, while the concentration of free MNPs decreases. This dynamic disrupts the initial saturation state, causing the RD profile to transition into an exponential decay mechanism. Interestingly, the RD plateau of mPVP1 and mPVP2 are notably lower [Figure 4J]. This can be attributed to extra PVP’s mechanism: it preferentially occupies adsorption sites on the ZIF-8 shell, which consequently depresses the maximum saturation plateau.

While the aggregation ratio metric can reflect the degree of MNPs dispersion, its “dichotomous” evaluation method is unable to reveal the continuous distribution characteristics of MNPs. To address this, we introduce NND analysis to provide a more refined statistical description of the spatial arrangement. The NND follows single-peaked Gaussian distribution, which reflects MNPs exhibit isotropic, random dispersion in 3D space [Supplementary Figure 9 and Figure 4K]. The fitting parameters summarized in Supplementary Table S8. The fitting curves for samples without extra PVP addition are remarkably consistent (peak: ~ 23 nm, full width at half maximum, FWHM: ~ 9 nm), which reflects that the MNPs in these samples possess highly similar interparticle distances and dispersion uniformity. These conditions do not affect the actual dispersion characteristics of the MNPs. In comparison, extra PVP addition samples exhibited concentration-dependent positive shifts in peak position (mPVP1: 25.03 nm, mPVP2: 36.57 nm) and concomitant broadening of FWHM (mPVP1: 16.59 nm, mPVP2: 32.60 nm). The positive shift in peak position reflects an increase in the average NND between MNPs. Given the constant MNPs size, this increased interparticle distance directly accounts for the observed reduction in the aggregation fraction. Furthermore, the concurrent broadening of FWHM indicates that the increase in NND is spatially heterogeneous. Free-state PVP molecules stochastically block ZIF-8 adsorption sites, thereby inducing a more disordered spatial distribution of MNPs. This contrasts with an alternative mode where PVP preferentially forms rigid steric barriers around MNPs - a scenario that would not produce significant FWHM broadening accompanying the peak shift.