Imaging thick objects with deep-subangstrom resolution and deep-subpicometer precision Authors: Wenfeng Yang, Haozhi Sha, Jizhe Cui, Rong Yu Published: 2026 ABSTRACT Size effects are ubiquitous in the structural, mechanical, and physical properties of materials, making it highly desirable to study the intrinsic properties of thick objects through high-resolution structural analysis in transmission electron microscopy. Although deep-subangstrom resolution has been achieved with multislice electron ptychography, the sample thickness is typically very limited. By combining energy filtering and extended local-orbital ptychography (eLOP) that retrieves varying aberrations during electron beam scanning, here, we report ptychographic reconstructions for silicon as thick as 85 nanometers, approximately three times larger than usual thickness threshold for conventional multislice electron ptychography. The elimination of aberration variations contributes to accurate reconstructions with an information limit of 18 picometers and atomic position precision of 0.39 picometers. Accurate ptychographic reconstructions for thick objects can facilitate the discovery or interpretation of intrinsic structural and physical phenomena in solids, which is of great importance in physics, chemistry, materials science, and semiconductor device engineering. INTRODUCTION Transmission electron microscopy (TEM) (1) is a widely used technique for real-space imaging of solids at high spatial resolution. With the development of ptychographic algorithms (2–7), TEM is increasingly used with ptychographic phase retrieval for ultrahigh spatial resolution and high phase accuracy. On a modern transmission electron microscope equipped with an aberration corrector (8, 9), the information limit of electron ptychography reached 39 pm in 2018 (10), 23 pm in 2021 (11), and 14 pm in 2023 (7). Recently, an information limit of 41 pm for two-dimensional (2D) materials was achieved using an uncorrected electron microscope (12). Deep-subangstrom-resolution (<0.5 Å) imaging with high phase accuracy helps to accurately characterize the microstructure of materials (13), e.g., orientation mapping around dislocation kinks in a metal oxide (14), lattice-resolved antiferromagnetic imaging (15), subangstrom resolution imaging of beam-sensitive materials (16–18), measuring atomic positions with subpicometer precision in a polar metal (19), measuring the concentration of oxygen vacancies in a superconductor (20), imaging polar vortex in twist SrTiO3 (21), and counting and positioning single atoms of interstitial oxygen in a metal (22). Atomic resolution imaging usually requires very thin samples, typically less than 10 nm for aberration-corrected TEM imaging and 20 nm for scanning TEM (STEM) imaging. High-quality ptychographic reconstructions also necessitate a sample thickness less than 20 nm for dense solids such as metals, ceramics, and semiconductors. The thickness threshold of electron ptychography of dense solids reached 21 nm in 2021 (11), 34 nm in 2024 (14), and 50 nm in 2024 (7). Size effects are pervasive in the structural, mechanical, and physical properties of materials. The reduction of sample thickness often results in altered structural or dynamical behaviors due to strain relaxation, surface diffusion and damage, or even phase transitions. Excessively thin samples pose substantial challenges for the reliability of structural analysis in TEM. To obtain intrinsic structures of solids, it is highly desirable to use samples that are as thick as possible for accurate structural analysis. In this study, we developed a method named extended local-orbital ptychography (eLOP) that enables the variation of aberrations as the electron beam scans across the sample. It should be noted that eLOP and its precursor (LOP) (7) are phase retrieval methods of unknown structures that starts from scratch. They differ from the method proposed in the reference (23), which involve the structural refinement of known initial models. By combining the eLOP method with energy-filtering, ptychographic reconstructions are realized for 85-nm-thick Si and 60-nm-thick SrTiO3, with the information limit of 18 pm for Si and 16 pm for SrTiO3. The atomic position precisions for Si and SrTiO3 are 0.39 and 0.42 pm, respectively, both reaching the deep-subpicometer (<0.5 pm) region. RESULTS Probe aberrations negatively affect high-resolution imaging of objects. In general, the problem is more severe for thick objects due to the larger focus range of the probe within the objects in the STEM mode. Conventional pixelated ptychography (CPP) describes the probe as pixel arrays and optimizes the probe to eliminate the influence of aberrations, which are assumed to be constant during probe scanning. However, aberrations often vary in experiments. eLOP describes the probe as position-dependent aberration functions, effectively eliminating the influence of aberration variations on ptychography (see Materials and Methods for details). An accurate probe description is essential for reconstructing thick samples and precisely determining atomic positions. Figure 1 shows the simulation tests conducted to evaluate the performance of the eLOP method in handling aberration variations, which are difficult to be considered in CPP (24). Figure 1 (A and B) shows a probe with the aberrations dependent on the scanning positions. The variations in the real part of the aberration coefficients A1 and A2 used in the simulations are shown in fig. S1 (A and B), respectively. Ptychographic reconstructions on simulated datasets with aberration variations. (A) Schematic illustration of the scanning probe with position-dependent aberrations. (B) Four representative probe amplitudes used in simulations, corresponding to the aberrations at the four corners shown in fig. S1 (A and B). a.u., arbitrary units. (C) The object phase reconstructed via CPP for a 20-nm-thick object. The atom in the inset is magnified four times. (D) The object phase reconstructed via eLOP for the 20-nm-thick object. The atom in the inset is magnified four times. (E) The object phase reconstructed via CPP for the 60-nm-thick object. (F) The object phase reconstructed via eLOP for a 60-nm-thick object. The atomic structure is overlaid on the phase images. Sr, Ti, and O atoms are shown with green, blue, and red circles, respectively. For a 20-nm-thick object, the object phase reconstructed via CPP is shown in Fig. 1C. The atom in the inset is magnified four times to illustrate the irregular atomic shape. Because the retrieved probe is independent of scanning positions, it can account only for the average aberrations over the scanned area. The residual aberrations in the probe are transferred into the object phase, resulting in the irregular atomic shapes and apparent lattice distortions as shown in fig. S2A. To quantitatively evaluate the precision of atomic positions, the distance (dSr-TiO) between the TiO column and Sr column was measured. Figure S3A shows the histogram of dSr-TiO with an SD of 1.12 pm. That means that CPP is not able to remove the effect of aberration variations, which is reflected in the object phase and in the convergent beam electron diffraction (CBED) pattern as shown in fig. S4. In contrast, eLOP provides highly accurate atomic shape, as shown in Fig. 1D, and precise atomic positions in the object phase, as shown in fig. S2B. The SD of dSr-TiO is only 0.11 pm as shown in fig. S3B, due to the removal of aberration variations as shown in fig. S1 (C and D). With increasing thickness, the influence of aberration variations on ptychographic reconstructions becomes stronger. For a 60-nm-thick object, CPP fails to reconstruct the object phase, as shown in Fig. 1E. In contrast, the object phase obtained by eLOP in Fig. 1F is accurate, with an SD of dSr-TiO measuring 0.28 pm, as shown in fig. S3C. The aberration maps in fig. S1 (E and F) closely match the true values used in the simulations. The precise reconstruction achieved by eLOP facilitates accurate measurement of the intrinsic physical behavior of materials. As shown in fig. S5, the reconstructed results of simulated BaTiO3 dataset are presented. The simulated dataset was generated using the aberrations illustrated in fig. S5A. As indicated in fig. S5B, similar to the case of SrTiO3, the eLOP method yields a highly accurate atomic shape. Figure S5C displays the polar displacement of Ti atoms with respect to the O atom centers (pTi-O). The eLOP method reliably provides both the magnitude and direction of pTi-O, whereas the CPP method, unable to eliminate residual aberrations, produces substantial variations in pTi-O across different regions, causing the polar displacements to deviate from the true values in the simulated structure. Simulated datasets were used to demonstrate the ability of the eLOP method for the reconstruction of thick objects. For comparison, reconstructions using the CPP method were also performed. The multislice method (25) as implemented in the μSTEM code (26) was used. The 4D-STEM datasets for various thicknesses were generated, and ptychographic reconstructions were carried out. Typical diffraction patterns are shown in fig. S6. Each diffraction pattern was padded to 256 × 256 pixels for better real-space sampling during the reconstructions. For the CPP method, we used the EMPTY program, which implements the multislice ptychography with adaptive propagator (6). The maximum thickness reconstructed by eLOP is at least 60 nm, whereas by CPP is 40 nm, as shown in Fig. 2 (A and B). The reason of larger thickness threshold of eLOP can be attributed to the robustness of the parameterized probe and object. The probe in CPP is composed of uncorrelated pixels. For 50- and 60-nm-thick objects, the probe cannot converge to a well-defined shape during the optimization. In contrast, the probe reconstructed via eLOP consistently maintains a well-defined shape, which benefits from the parameterization of the probe. Additionally, the object in eLOP is more robust than that in CPP. Although a 40-nm-thick object can be reconstructed using CPP, the sample thickness must be known a priori and fixed during the reconstruction; otherwise, the method diverges, resulting in unsuccessful reconstruction. In contrast, eLOP allows the sample thickness to be optimized during reconstruction. This capability is highly advantageous, as the sample thickness is usually unknown in experiments. Comparison of CPP and eLOP for different thicknesses on simulated datasets of SrTiO3. (A and B) Object phase reconstructed via the CPP method for different object thicknesses. The object thicknesses are indicated at the top. The atomic structure is overlaid on the phase image. Sr, Ti, and O atoms are shown with green, blue, and red circles, respectively. Scale bars, 2 Å. A SrTiO3 sample with a thickness gradient was used to demonstrate the performance of the eLOP method in enhancing the thickness threshold. Figure 3A shows a low-magnification image of the sample. The thickness of the sample in different areas was estimated using low-loss electron energy loss spectroscopy (EELS), with the results shown in Fig. 3 (B and C). Through the comparison between the true thickness derived from ptychographic reconstruction and the relative thickness (a value proportional to the inelastic mean free path, λ) estimated by EELS, the inelastic mean free path (λ) of SrTiO3 was measured to be 115 nm. The elemental mapping is shown in fig. S7. Energy-filtered 4D-STEM datasets used for ptychographic reconstructions were collected from the four areas indicated in the low-magnification image, as shown in Fig. 3D. Typical diffraction patterns are shown in figs. S8 and S9. Each diffraction pattern was padded to 256 × 256 pixels for fine real-space sampling during ptychographic reconstruction. Energy-filtering 4D-STEM experiment. (A) Low-magnification image of the SrTiO3 specimen used for acquiring the experimental datasets. Four areas with different thicknesses selected for data collection are marked. Scale bar, 20 nm. (B) Thickness map obtained from low-loss electron energy loss spectroscopy (EELS). (C) Thickness variation along the profile in (B), indicating that areas 1 to 4 have approximate thicknesses of 30, 40, 50, and 60 nm, respectively. (D) Schematic illustration of the energy-filtering 4D-STEM experiment for ptychography. Figure 4 shows the reconstructed phase images of different areas via different methods. The electron doses for regions of different thicknesses are listed in table S1. The slice thickness is 1 nm in the reconstructions. For the 30-nm-thick area, CPP shows good convergence for the energy-filtered dataset (EF-CPP) but fails to converge for the unfiltered dataset (UF-CPP), highlighting the beneficial effect of removing inelastic scattering in ptychographic reconstructions. For thicker areas, CPP fails to converge for both filtered and unfiltered datasets. In contrast, eLOP exhibits good convergence for the unfiltered datasets (UF-eLOP) in areas of up to 50-nm thickness and for the filtered datasets (EF-eLOP) in areas of up to 60-nm thickness. The improvement in the achievable thickness threshold of eLOP compared with CPP is therefore substantial. The reconstruction videos, illustrating the phase evolution of the object function over the first 200 iterations of eLOP for 50- and 60-nm-thick specimens, have been deposited on Zenodo (27). Comparison of CPP and eLOP at different sample thickness for unfiltered and filtered diffraction patterns. The sample thicknesses are indicated at the top of each panel. The atomic structure is overlaid on the phase image. Sr, Ti, and O atoms are shown with green, blue, and red circles, respectively. Scale bar, 2 Å. For the filtered dataset in the 60-nm-thick area, eLOP also fails to converge when the position dependence of the probe is disabled, indicating the necessity of incorporating the position-dependent probe aberrations for thick objects. Comparison of single CBED pattern is shown in fig. S10. The variation of aberrations is shown in fig. S11. The A2 and B2 aberrations remain nearly constant, whereas A1 exhibits substantial changes. Figure 5 shows the information limit and the atom-position precision in ptychographic phase images of thick objects. The diffraction patterns are padded to 420 × 420 pixels for fine real-space sampling during reconstructions. Figure 5A shows the phase image averaged over the 60 slices of the 60-nm-thick SrTiO3 specimen. The information limit reaches 16 pm, as shown in Fig. 5B. The distance dSr-TiO between the Sr and TiO atomic columns is 272.2 pm, with an SD of 0.42 pm, reaching the deep-subpicometer regime. For individual slices, the SD of dSr-TiO range from ~0.6 to 1.1 pm (fig. S12), indicating that thicker objects provide a higher signal-to-noise ratio and are beneficial for achieving high precision in determining atomic positions. Information limit and atom-position precision in ptychographic phase images of thick objects. (A) The object phase of the 60-nm-thick SrTiO3 along the [001] zone axis. Scale bar, 2 Å. (B) The diffractogram corresponding to (A). Scale bar, 2 Å−1. (C) The histogram of dSr-TiO with an SD of 0.42 pm. (D) The object phase of the 85-nm-thick Si along the [001] zone axis. The atomic structure is overlaid on the phase image. Scale bar, 2 Å. (E) The diffractogram corresponding to (D). Scale bar, 2 Å−1. (F) The histogram of dSi-Si with an SD of 0.39 pm. a.u., arbitrary units. Silicon, as the most important semiconductor, is used to further evaluate the thickness threshold of eLOP. Typical experimental CBED patterns and the position-averaged CBED patterns are shown in fig. S13. Figure 5D shows the reconstructed phase image of an 85-nm-thick Si specimen along the [001] zone axis. The information limit reaches 18 pm, as shown in Fig. 5E. The distance dSi-Si between neighboring Si atomic columns is measured to be 190.6 pm, with an SD of 0.39 pm, also reaching the deep-subpicometer regime. The eLOP method provides a powerful tool for the precise characterization of intrinsic physical behavior in thick objects. The results for BaTiO3 reconstructed via eLOP are shown in fig. S14. The presence of the heavy element Ba and the strong dynamical diffraction in BaTiO3 pose substantial challenges for structure reconstruction of the relatively thick sample (25 nm) using the CPP method. The average phase of the object function and the phase at the top surface obtained by the eLOP method are displayed in fig. S14 (A and B), respectively. Surface defects are observed in the blue-boxed region of the top layer as shown in fig. S14B, inducing the variation of pTi-O along the depth direction within the defect region. As shown in fig. S14C, large atomic displacements are observed in the vicinity of the defect near the top surface, pointing downward and to the right. As the defects disappear in the bulk layers, the displacements become smaller and predominantly oriented upward. DISCUSSION By combining eLOP with energy filtering, deep-subangstrom resolution and deep-subpicometer precision have been achieved for dense solids as thick as 85 nm. This method alleviates the demanding requirements of electron ptychography on sample thickness, paving the way for ultrahigh-resolution imaging and accurate analysis of intrinsic structures and phenomena in engineering materials and semiconductor devices. MATERIALS AND METHODS Details of eLOP Because the aberration of the probe will change during the experiments, the variation of aberration coefficients at different scanning positions needs to be considered in the LOP algorithm. In this work, the probe at the jth position is expressed asPj(r)=FA(k)eiχ(k,j)(1)where A(k) is the aperture function. The scanning position–dependent aberration function χ(k,j) is written as (8, 28, 29)χ(k,j)=ℝ12ωω¯C1(j)+12ω¯2A1(j)+ω2ω¯B2(j)+13ω¯3A2(j)+14ω2ω¯2C3(j)+…(2)where ℝ{·} means taking the real part of the matrix in the bracket. ω=kx+i∗ky, where kx and ky are coordinates in reciprocal space, respectively. ω¯ is the complex conjugate matrix of ω. C1(j),A1(j),B2(j),A2(j), and C3(j) are the mapping of aberration coefficients. To account for the noncircular aperture shape, an elliptical aperture was used in this study. Additionally, acquiring a diffraction pattern of the vacuum region helps obtain a more accurate aperture shape. Materials and sample preparation The SrTiO3, Si, and BaTiO3 single crystals were purchased from Hefei Kejing Materials Technology Co. Ltd. The STEM samples were prepared using the standard lift-out method by focus ion beam (FIB) in a Zeiss Auriga FIB instrument. Experimental details and data processing EELS experiments were carried out using an FEI Titan3 Themis 60-300 instrument. The spectrum image data and the annular dark-field signal were acquired with a beam current of about 50 pA, a convergence semiangle of 25 mrad. The measured thickness-to-λ ratio of area 1 in Fig. 3A is 0.26 nm−1. Given that the reconstructed thickness is 30 nm, the used inelastic mean free path of SrTiO3 is 115 nm. The 4D-STEM diffraction datasets were acquired using an electron microscope pixel array detector (K3) installed on a probe aberration–corrected JEM-ARM300F2 GRAND ARM2 atomic resolution analytical microscope operated at 300 kV. An energy-selecting slit with a width of 5 eV was inserted with the zero loss peak at the center to exclude most inelastically scattered electrons in diffraction patterns. The convergence semiangle was set to 24.3 mrad. Each diffraction pattern has a dimension of 128 × 128, and the camera length is 1.5 cm, giving a reciprocal pixel size of 0.035 Å−1 (0.68 mrad). To achieve substantial overlap, the defocus and scanning step size of both SrTiO3 and silicon are 30 and 0.03 nm, respectively. The thickness of SrTiO3 is 30 to 60 nm. The thickness of silicon is 85 nm. The defocus, thickness, scanning step size, and reciprocal pixel size of BaTiO3 dataset are 25 nm, 25 nm, 0.04 nm, and 0.031 Å−1, respectively. All data that support the findings of this study have been deposited in Zenodo (30). Multislice simulation of 4D datasets μSTEM is used to simulate 4D datasets for ptychographic reconstruction. The lattice constant of SrTiO3 was set to 0.4 nm. For the simulated datasets used for Fig. 1, the probe was formed with an acceleration voltage of 300 kV, and a convergence angle of 24.3 mrad. Each diffraction pattern has a dimension of 128 × 128 pixels. The probe was formed with a slight defocus of 20 nm and a step size of 0.04 nm for SrTiO3 with the thickness of 20 nm, while a slight defocus of 30 nm and a step size of 0.03 nm for SrTiO3 with a thickness of 60 nm to obtain a bigger overlap. For the simulated datasets used for Fig. 2, the probe was formed with an acceleration voltage of 300 kV, a slight defocus of 30 nm, a step size of 0.03 nm, and a convergence angle of 24.3 mrad. Each diffraction pattern has a dimension of 128 × 128 pixels. The sample thicknesses were set to 30, 40, 50, and 60 nm. For the simulated dataset used for fig. S5, the probe was formed with an acceleration voltage of 300 kV, a slight defocus of 20 nm, a step size of 0.04 nm, and a convergence angle of 24.3 mrad. Each diffraction pattern has a dimension of 128 × 128 pixels. The sample thickness was set to 20 nm.