Data from: Point defects and their local environments in Ce1-xZrxO2 as revealed by multislice electron ptychography
Data files
Oct 22, 2026 version files 19.64 GB
-
README.md
27.76 KB
-
ZDCData_Dryad.zip
19.64 GB
Abstract
This dataset contains the experimental four-dimensional scanning transmission electron microscopy (4D-STEM) data, multislice electron ptychography (MEP) reconstructions, defect analysis results, multislice simulations, and trained machine-learning classifier supporting the study “Point defects and their local environments in Ce1-xZrxO2 as revealed by multislice electron ptychography.” Epitaxial Ce1-xZrxO2 thin films (x = 0, 0.02, and 0.1) were grown by pulsed-laser deposition on (001) yttria-stabilized zirconia and prepared as cross-sectional lamellae by focused ion beam milling. Raw 4D-STEM datasets were acquired along the [110] zone axis on an aberration-corrected Thermo Fisher Themis Z at 300 kV using an EMPAD detector (27 mrad semi-convergence angle, 0.53 Å scan step, approximately 3 × 105 e/Å2), with three acquisitions each for CeO2 and Ce0.9Zr0.1O2 and one for Ce0.98Zr0.02O2. Each acquisition is provided with its metadata, scan calibration, and overview image. Depth-resolved reconstructions were performed with the open-source Python package phaser, and the reconstruction recipe is included. For each dataset, the deposit provides the reconstructed object and probe, refined scan positions and specimen tilts, aligned phase stacks, atomic-column statistics, interstitial detection and depth-profile metrics, and final interstitial classifications. The Ce0.9Zr0.1O2 data also include a series of reconstructions run with different Poisson noise-model parameters. Simulated 4D-STEM datasets matched to the experimental conditions are provided for CeO2 supercells containing one, two, or three interstitial Zr atoms, an interstitial Ce, an interstitial Ga, a three-atom interstitial O cluster, and interstitial Zr with neighboring Ce displaced by 0.1 to 0.3 Å, together with their MEP reconstructions. The random-forest classifier trained on these simulations to distinguish isolated from overlapping interstitial Zr is included with its validation results. Together, these data support the identification of Zr at octahedral interstitial sites, a concentration of approximately 1020 cm-3, and the associated lattice expansion, oxygen-vacancy coupling, and Zr–O bonding disorder. They can also be reused to benchmark ptychographic reconstruction algorithms and to develop point-defect detection methods.
Authors: Jingrui Wei*, Sooyoon Hwang*, Zouhair Hanani, Andrey Shults, Jake A. DeChiara, Jon-Paul Maria, Lane W. Martin, James M. LeBeau (* equal contribution)
This dataset contains the experimental four-dimensional scanning transmission electron microscopy (4D-STEM) data, multislice electron ptychography (MEP) reconstructions, interstitial-defect analysis results, multislice simulations, and the trained random-forest classifier used in the associated publication.
The aim of the study is to locate individual cation interstitials (primarily Zr interstitials, Zri) in three dimensions inside Ce1−xZrxO2 thin films and to identify how many atoms occupy each interstitial site. Films with x = 0 (CeO2), x = 0.02 and x = 0.1 were imaged along the [110] zone axis at 300 kV with an electron microscope pixel array detector (EMPAD). Each 4D-STEM dataset was reconstructed with the multislice ptychography code phaser, which yields the projected electrostatic potential (as a phase shift) of the specimen in a stack of thin depth slices. Bright features at the empty octahedral interstitial sites of the fluorite lattice were detected in these slices, tracked along depth, and characterized by their integrated phase and depth width. These metrics were compared with reconstructions of multislice simulations containing known interstitial configurations, and a random-forest classifier trained on the simulations was used to assign each experimental interstitial to a class (1Zr, 2Zr, 3Zr, or weaker than a single Zr).
Total size is approximately 21 GB.
Software needed
- Python 3.10 or newer with NumPy ≥ 2.0 (required to read the
.npz,.npyand.pklfiles), h5py (.h5), pandas ≥ 3.0 (DataFrames inside.pklfiles; some were written with pandas 3 and fail to load with pandas 2.x), and scikit-learn ≥ 1.7 (the classifier). - phaser (https://github.com/hexane360/phaser) to rerun reconstructions from the raw data with the recipe file. The
.h5reconstruction files can be read with h5py alone. - Any TIFF viewer that supports multi-page 32-bit images (for example ImageJ/Fiji, or the Python package tifffile) for
.tif/.tifffiles. - Any web browser for
.svgfiles, and any PDF viewer for.pdffiles. - The raw
.rawfiles are flat binary files (see "Raw 4D-STEM data" below) and can be read withnumpy.fromfile, or with py4DSTEM or similar packages that support the EMPAD format.
Conventions used throughout
- Lengths are in ångström (Å) unless the key or column name ends in
_nm(nanometres) or_px(pixels). Angles are in milliradians (mrad). - Reconstruction pixel size is 0.1493 Å in both x and y for all reconstructions. Depth slices are 3.8 Å thick; slice index k is at depth 3.8·k Å from the top of the reconstruction volume.
- "Phase" is the phase shift (radians) imparted by each depth slice of the reconstructed object, which is proportional to the projected electrostatic potential of that slice. "Integrated potential" or "potential" in the analysis files means phase summed over the pixels inside a disk of a given radius around an atomic column or interstitial site (and, for depth profiles, along depth). These values are in arbitrary units and are directly comparable only between files produced by the same pipeline (experiment vs. simulation in this dataset).
- Folder and file names use
das a decimal point, so0d4= 0.4,0d2A= 0.2 Å,1d0= 1.0. eps(ε) is the regularization parameter of the Poisson noise model used in the phaser reconstruction.eps0d4(ε = 0.4) is the value used for all final results.iterNin a file name gives the total number of reconstruction iterations at which that output was saved (for exampleiter256.h5).fp0d2Ain simulation folder names denotes the frozen-phonon root-mean-square displacement used in the multislice simulation (0.2 Å).- Sample sets correspond to the three film compositions. The datasets labelled A, B and C for each composition match datasets A, B and C in Supplementary Table 3 of the article.
- File paths written inside metadata and recipe files refer to the original computing environment and do not need to exist for the data to be used.
- Not every analysis folder contains every file type listed below. Folders processed earlier in the project contain a subset of the outputs.
Folder structure
Compressed directory: ZDCData_Dryad.zip
Data/
├── README.md
├── Experiment/
│ ├── exp_phaser_recipe.yaml
│ ├── CeO2/ # x = 0
│ │ ├── acq4/ acq6/ acq7/ # raw 4D-STEM data, datasets A, B, C
│ │ └── recons_analysis/
│ │ ├── acq4_11/ # reconstruction + analysis of acq4
│ │ ├── acq6_01/ # ... of acq6
│ │ └── acq7_07/ # ... of acq7
│ ├── Ce0.98Zr0.02O2/ # x = 0.02
│ │ ├── acquisition_3/ # raw 4D-STEM data
│ │ └── recons_analysis/
│ │ └── acq3_03/
│ └── Ce0.9Zr0.1O2/ # x = 0.1
│ ├── acq1/ acq3/ acq7/ # raw 4D-STEM data, datasets A, B, C
│ └── recons_analysis/
│ ├── acq1_eps0d4_afterinit/ # final reconstruction, dataset A (first 128 × 128 scan half)
│ ├── acq1_bulkhalf/ # dataset A, second 128 × 128 scan half
│ ├── acq3_eps0d4/ # final reconstruction, dataset B
│ ├── acq7_eps0d4_afterinit/ # final reconstruction, dataset C
│ └── acq7_eps{0d1,0d6,0d8,1d0}_afterinit/ # dataset C reconstructed with other ε values
└── Simulation/
├── multislice/ # simulated 4D-STEM datasets, one folder per defect configuration
├── phaser/
│ ├── element_series/ # MEP of each simulated configuration (ε = 0.4)
│ ├── eps_series/ # MEP of the 1Zr simulation with other ε values
│ └── Ce_shift_series/ # MEP of Zr_i with displaced neighbouring Ce
└── RFclassifier/ # trained random-forest classifier
Naming of the experimental reconstruction folders is acquisition_suffix. For CeO2 and x = 0.02 the numeric suffix (_11, _01, _07, _03) is an internal reconstruction run number. For x = 0.1 the suffix gives the noise-model parameter (eps0d4 etc.), and afterinit means the reconstruction was started from a per-position tilt map tracked from the raw data (see init: tilt in the recipe). In diagnostic image names, sampleNN (for example sample17) is the internal laboratory sample number of the film.
1. Raw experimental 4D-STEM data (Experiment/<composition>/acq*/, acquisition_3/)
Each acquisition folder holds one 4D-STEM scan and contains five files.
| File | Description |
|---|---|
scan_x<Nx>_y<Ny>.raw |
4D-STEM dataset from the EMPAD detector. Flat little-endian float32 binary with no header. Shape is (Ny, Nx, 130, 128), i.e. one 128 × 130 pixel diffraction pattern per probe position, with the last two detector rows containing frame metadata (discard them to obtain 128 × 128 patterns). Values are detector counts in ADU after background subtraction; divide by adu in the JSON (867 ADU per electron) to convert to electrons. |
<name>.json |
Acquisition metadata (see keys below). |
<name>_calib.json |
Same as <name>.json with the measured scan-distortion correction added in scan_correction. This is the file read by phaser. |
<name>.xml |
Log written by the EMPAD acquisition software (timestamps, software version, camera and scan settings, background image used). |
scan_roi3_cw_annulus_x64_y64_ri20_ro50.tif |
Virtual annular image computed from the 4D-STEM data at acquisition (annulus centred at detector pixel (64, 64), inner radius 20 px, outer radius 50 px), for orientation only. |
Scan sizes are 150 × 300 (CeO2 acq4), 150 × 150 (CeO2 acq6, acq7 and x = 0.02 acquisition_3) and 128 × 256 (all x = 0.1 acquisitions) probe positions.
Keys in the acquisition JSON files (SI units)
voltageaccelerating voltage (V), 300 kVconv_angleprobe convergence semi-angle (mrad), 27 mraddefocusprobe defocus (m)camera_length(m)diff_stepangular size of one detector pixel (mrad per pixel)scan_shapenumber of probe positions [Ny, Nx]scan_stepprobe step size [y, x] (m)scan_fovscan field of view (m)scan_rotationscan rotation (degrees)exposure_timedwell time per probe position (s)beam_currentprobe current (A)adudetector counts per incident electronscan_correction2 × 2 affine matrix correcting scan distortion (only in_calib.json)time,time_unix,bg_unix,has_bg,orig_path,author,notes,crop,scan_positionsbookkeeping fields
Experiment/exp_phaser_recipe.yaml is the phaser input file used for the final x = 0.1 reconstructions (shown for dataset A). It defines the raw data input, the real-space crop, the initial tilt map, the number and thickness of depth slices, and four successive gradient-descent engines with their learning rates, regularization weights, noise model (Poisson, ε) and constraints. Reconstructions of the other datasets used the same recipe with the data path, crop, ε and iteration counts changed as indicated by the folder names.
2. Reconstructions and analysis (Experiment/<composition>/recons_analysis/<dataset>/)
All reconstruction folders share the same set of file types. The folder Ce0.9Zr0.1O2/recons_analysis/acq1_eps0d4_afterinit/ contains the complete set and is described here as the example. The same structure applies to all other folders under recons_analysis/ and under Simulation/phaser/.
2.1 Reconstruction output
iterN.h5 phaser reconstruction state saved after N total iterations (HDF5). Objects inside
| Path | Shape, type | Description |
|---|---|---|
object/data |
(n_slices, Ny, Nx) complex64 | Reconstructed complex transmission function of each depth slice. np.angle() gives the phase (rad), np.abs() the amplitude. Dimension labels z, y, x. |
object/sampling |
(2,) float64 | Pixel size [y, x] (Å). |
object/extent |
(2,) float64 | Size of the object array [y, x] (Å). |
object/corner |
(2,) float64 | Real-space coordinate of pixel (0, 0) [y, x] (Å). |
object/region_min, object/region_max |
(2,) float64 | Bounding box of the scanned region [y, x] (Å). |
object/zs |
(n_slices,) float32 | Depth of each slice (Å). |
object/thicknesses |
(n_slices,) float32 | Thickness of each slice (Å), 3.8 Å. |
probe/data |
(n_modes, Ny, Nx) complex64 | Reconstructed mixed-state probe, 8 incoherent modes, real space. |
probe/sampling, probe/extent |
(2,) float64 | Probe pixel size and extent [y, x] (Å). |
scan |
(n_positions, 2) float64 | Refined probe positions [y, x] (Å). |
tilt |
(n_positions, 2) float64 | Refined local beam/specimen tilt at each probe position [y, x] (mrad). Experimental reconstructions only. |
wavelength |
scalar float64 | Electron wavelength (Å). |
iter/total_iter, iter/engine_num, iter/engine_iter |
scalar | Total iteration count, index of the engine, and iteration within that engine at save time. |
progress/<metric>/values, progress/<metric>/iters |
(n,) | Convergence history. Metrics are detector_loss and total_loss (cost function), obj_l1, obj_l2, obj_tikh, probe_recip_tikh (regularization terms), pos_update_rms (Å) and tilt_update_rms (mrad) (RMS position and tilt updates), time and utc (wall-clock time, s). |
type, version |
string | File type (phaser_state) and format version. |
Images exported by phaser (32-bit TIFF, pixel size 0.1493 Å, same iteration as the .h5)
object_phase_stack_iterN.tiffphase (rad) of every depth slice, one page per sliceobject_phase_sum_iterN.tiffphase summed over all slices (projected phase, rad)object_mag_stack_iterN.tiff,object_mag_sum_iterN.tiffthe same for the object amplitudeprobe_iterN.tiffreal-space intensity of the probe modesprobe_recip_iterN.tiffreciprocal-space (diffraction-plane) intensity of the probe modesscan_iterN.svgplot of refined scan positionstilt_iterN.svgmap of refined local tilts
2.2 Analysis input
phase_stack.npy (and identical phase_stack.tiff where present) float32 array (n_slices, Ny, Nx). The object phase (rad) of each depth slice after cropping to the scanned region and an affine correction and rotation that aligns the fluorite lattice with the image axes. All pixel coordinates in the analysis files below refer to this array. Pixel size 0.1493 Å, slice thickness 3.8 Å.
data4peakreview.pkl dict holding the inputs used to review depth-profile peaks (larger datasets only)
obj_p_affined_rotatedthe aligned phase stack (same asphase_stack.npy)position_matrix(n_columns, 2) pixel coordinates of each tracked interstitial columnz_files_fulllist, one entry per tracked column, each a tuple of the two depth profiles (integrated potential vs. slice, for the two integration radii)RADII_1,RADII_2the two integration radii (Å), 0.15 and 0.3slice_cut,extend_zsslice range analysed and the number of extra slices added at each end of the depth profilesfpathoriginal path of the source.h5file
2.3 Atomic-column statistics
cation_anion_stats_zblur.npz (named cation_anion_stats_blur_fit.npz or acq3_cation_anion_stats_blur_fit.npz in some folders, same keys). Positions and integrated phase of every lattice cation (Ce/Zr) and anion (O) column in each analysed slice. In "zblur" files the stack was smoothed slightly along depth before integration.
| Key | Shape | Description |
|---|---|---|
final_cation_pos |
(n_slices,) object array of (n_cations, 2) | Fitted cation column positions in each slice (pixels, in the phase_stack frame). |
final_cation_int |
(n_slices,) object array of (n_cations,) | Phase integrated inside a disk of radius cation_radii around each cation (arbitrary units). |
final_anion_pos, final_anion_int |
as above | The same for anion columns, radius anion_radii. |
px_size_angstrom |
scalar | Pixel size (Å). |
cation_radii, anion_radii |
scalar | Integration radii (Å). |
slice_cut |
(2,) int | First and last slice index analysed. Entry i of the per-slice arrays corresponds to slice slice_cut[0] + i. |
surface_ss |
(2,) int | Slice indices identified as the top and bottom specimen surfaces. |
In simulation folders the per-slice arrays are regular 3D arrays (n_slices, n_columns, 2) instead of object arrays because the number of columns does not change.
2.4 Interstitial detection and depth profiles
interstitial_stats.npz (named interstitial_stats_ccthre0.3.npz in simulation folders, where ccthre0.3 is the detection cross-correlation threshold of 0.3 used).
| Key | Shape | Description |
|---|---|---|
raw_full_pos_px |
(n_slices,) object array of (n, 2) | Positions (pixels) of the intensity maxima detected at octahedral interstitial sites in each slice before depth tracking. |
raw_full_int_radii2 |
(n_slices,) object array of (n,) | Phase integrated within the larger radius (radiis[1]) at each detected position (arbitrary units). |
tracked_z_potentials |
(n_columns, 2, n_z) float64 | Depth profile of integrated phase for each interstitial column after tracking through depth. Axis 1 indexes the two integration radii in radiis. Axis 2 runs over slices z_slice_range[0] to z_slice_range[1]. |
radiis |
(2,) | Integration radii (Å), 0.15 and 0.3. |
slice_cut, surface_ss |
(2,) int | As in 2.3. |
z_slice_range |
(2,) int | Slice range covered by the depth profiles. |
z_profile_peaks_analysis.pkl pandas DataFrame with one row per peak found in each depth profile. The row index is a MultiIndex of (column number as a two-digit string, profile p1 for radius 0.15 Å or p2 for radius 0.3 Å, peak number). Columns
peak_xdepth of the peak maximum (nm)peak_yintegrated potential at the peak (arbitrary units)fwhmfull width at half maximum of the peak along depth (nm)fwfhfull width of the peak at a fixed height (50 and 100 V*Å3 for radii 0.15 Å and 0.3 Å respectively) (nm), a measure of the depth extent of the peak tailsfwhm_estimated,fwfh_estimatedwidths estimated by extrapolation when the measured width is truncated by the end of the profile or by an overlapping peak (nm, NaN or None when not needed)left_x,right_xdepth limits of the peak (nm)heightreference level parameter used by the peak searchmethodwidth-measurement method (interpolation_upsampled)incomplete,overlap_affected,left_limited,right_limited,left_incomplete,right_incompleteboolean quality flags indicating truncated peaks or interference from neighbouring peaksraw_idxindex of the column intracked_z_potentials(present in later files)
inters_all_in_one.pkl Final result for the dataset. A dict of equal-length arrays/lists with one entry per identified interstitial
| Key | Description |
|---|---|
column_idx |
Index of the interstitial column (first axis of tracked_z_potentials). A column can host more than one interstitial at different depths. |
pos_xy_nm |
Lateral position (x, y) in the phase_stack frame (nm). |
pos_z_nm |
Depth of the interstitial (nm). |
potential1, potential2 |
Peak integrated potential with the 0.15 Å and 0.3 Å radius (arbitrary units). |
z_peak_fwhm_nm, z_peak_fwfh_nm |
Depth widths of the peak (nm), as defined above. |
d_from_surface |
Distance from the nearest specimen surface (nm). |
predicted_class |
Class assigned by the random-forest classifier, one of 1Zr, 2Zr, 3Zr, weaker (contrast weaker than a single Zr interstitial). |
all_probabilities |
dict of the classifier probability for each class. |
confidence |
Probability of the predicted class. |
margin |
Difference between the two highest class probabilities. |
top_2_classes |
The two most probable classes. |
is_confident, confidence_level |
Whether the assignment passes the confidence criterion, and the level (high, medium, low). |
Ce0.98Zr0.02O2/recons_analysis/acq3_03/interstitial_info_allinone.pkl is the same result in an earlier format. Its keys are column_idx, pos_xy_nm, pos_z_nm, potential_0d2A and potential_0d4A (integrated potential with 0.2 Å and 0.4 Å radii), id (assigned class), id_conf (class probability), z_peak_fwhm_nm and z_peak_fwfh_nm.
2.5 Diagnostic figures
z_profileN.pdfdepth profile of tracked column N with detected peaks and widths markedsampleNN[_acqK]_int_profiles_M.pngpanels of all depth profiles in the dataset (split over several pages M)confi_depth_section_intN_log.png,depth_section_intN_log.png,depth_section_NoN.pngdepth cross-sections (x–z) through interstitial N, log-scaled phase.confi_marks interstitials with a confident class assignmentlog_slice_slice_N.pnglog-scaled phase image of slice N;log_slice_10-23_animation.gifanimation through slices 10 to 23;log_slice_checked_inters_mean.pngmean of the slices containing reviewed interstitialscontrast_slices.pngcomparison of contrast across slicestotal_inters_map.png,total_inters_map_precheck.pdfmap of all detected interstitial positions (precheck means before manual review)
3. Multislice simulations (Simulation/multislice/<configuration>/)
Simulated 4D-STEM datasets of a 15 nm thick CeO2 supercell viewed along [011] (equivalent to [110]) containing a known arrangement of interstitial atoms. Each folder contains
scan_x80_y80.rawsimulated 4D-STEM data in the same EMPAD binary layout as the experiment (float32, shape (80, 80, 130, 128)), with Poisson noise for the dose given in the JSON<configuration>_scanmatch.json(or_sepa.json,Zri_displacedCe*.json) simulation metadata with the same keys as the experimental JSON plusdose(electrons per Å2),thickness_nm,zone(zone axis),aberration_A(residual aberrations),tilt,zr_percentandstructure_file(name of the atomic model used). Simulation parameters are 300 kV, 27 mrad convergence, 13 nm defocus, 0.53 Å scan step and 4.24 nm field of view, matching the experiment<configuration>.pngpreview image
Configurations (in the element configurations the supercell contains 36 interstitial columns with the defect placed at four different depths; the displaced-Ce configurations contain a single interstitial column with the defect at four depths)
1Zr,2Zr,3Zrone, two or three Zr atoms at octahedral interstitial sites in the same column (2Zruses a larger separation between the two Zr atoms, hence_sepa)1Ceone Ce interstitial1Gaone Ga interstitial, as a lighter-cation reference3Othree O interstitials in the same columnZri_disCe_0d1,Zri_disCe_0d2A,Zri_disCe_0d3one Zr interstitial with the neighbouring Ce atoms displaced away from it by 0.1, 0.2 and 0.3 Å
4. Reconstructions of simulations (Simulation/phaser/)
Folder names are <configuration>_eps<ε>_fp<frozen-phonon displacement>. The files and keys are identical to those described in section 2 for experimental reconstructions, except that the .h5 files contain no tilt dataset.
element_series/reconstructions of each configuration in section 3 with ε = 0.4eps_series/reconstructions of the1Zrsimulation with ε = 0.1, 0.6, 0.8 and 1.0, used to test the effect of the noise-model parameterCe_shift_series/Zr_disCe_0d{1,2,3}A/reconstructions of the displaced-Ce configurations
interstitial_fingerprint.pkl (simulation folders only). The metric "fingerprint" of the known interstitials, used as training data for the classifier. dict with potential_0d15A, potential_0d30A (peak integrated potential with 0.15 and 0.30 Å radii, arbitrary units), z_peak_fwhm_nm, z_peak_fwfh_nm (nm), one entry per interstitial (36 columns × 4 depths = 144 entries for the element series). The keys column_idx, pos_xy_nm, pos_z_nm, id, id_conf are present for format compatibility and are empty.
5. Classifier (Simulation/RFclassifier/)
rf_classifier.pkltrained scikit-learnRandomForestClassifier(400 trees, maximum depth 20,random_state=42). It takes four features per interstitial, the integrated potentials with the 0.15 Å and 0.3 Å radii and the two depth widths (FWHM and FWFH, nm), and returns the classes1Zr,2Zr,3Zr,weaker. Load withpickle.loadand callpredictorpredict_proba.confusion_matrix_rf.pngconfusion matrix of the classifier on held-out simulated datarf_figure_importance.pngrelative importance of the four input features
Sharing and access
The data are released under the CC0 license applied by Dryad. Please cite the associated article when using the data.
