A 4D-STEM Tomographic Framework Assisted by Object Tracking for Nanoparticle Structure Solution

Three-dimensional electron diffraction (3D ED) has emerged as a powerful method for solving the structures of sub-micron-sized particles down to nanoparticles. However, it faces technical challenges when applied to beam-sensitive samples or conglomerated and agglomerated nanoparticles. This study presents a novel approach that combines 4D-STEM tomography with object tracking and segmentation algorithms to overcome these limitations and achieve single-crystalline 3D ED datasets from nanopowder samples. The method and data quality are assessed on brookite TiO2 nanorods and beam-sensitive CsPbBr3 nanoparticles. To finely sample the reciprocal-space, the data acquisition was automated to acquire hundreds of 4D-STEM scans using a slightly convergent beam and at fine tilt steps. The proposed method provides enhanced signal-to-noise ratio, low illumination time for reducing beam damage, and the ability to analyze multiple particles from a single tomographic dataset. The procedure is optimized to be feasible using commercially available desktops and detectors. This extends the method applicability in the community to the systems and samples that were previously inaccessible for conventional 3D ED methods, particularly breaking the technical challenges for the data acquisition.

💡 Research Summary

This paper introduces a comprehensive 4D‑STEM tomography framework that integrates object tracking and segmentation to overcome longstanding challenges in three‑dimensional electron diffraction (3D ED) of nanomaterials. Traditional 3D ED requires a single, isolated crystal and suffers from drift, contamination, and beam‑damage, especially for beam‑sensitive particles or densely packed/agglomerated samples. The authors address these issues by employing a slightly convergent electron probe (semi‑angle 0.6–1.2 mrad) and very fine tilt increments (0.1°–0.25°) to acquire hundreds of 4D‑STEM scans automatically. Each scan records a diffraction pattern at every probe position, enabling the reconstruction of virtual images that reveal real‑space information with nanometer resolution.

In post‑processing, the virtual images are fed into two state‑of‑the‑art tracking algorithms: the Discriminative Correlation Filter with Channel and Spatial Reliability (CSRT) from OpenCV, and the Segment‑Anything‑Model 2 (SAM2) from Meta AI. These tools automatically locate regions of interest (ROIs) corresponding to individual particles, even when particles are only 20–30 nm apart or embedded in agglomerates. After ROI segmentation, diffraction patterns belonging to the selected region are summed, effectively discarding background contributions from the support membrane and dramatically improving signal‑to‑noise ratio (SNR).

The workflow consists of five steps: (1) automated acquisition of a tilt series of 4D‑STEM datasets, (2) generation of virtual images for visual inspection, (3) object tracking and segmentation to define ROIs, (4) extraction of 3D ED frames by summing diffraction patterns within each ROI, and (5) conventional data reduction, structure solution, and refinement using Jana2020, SUPERFLIP, and SHELXT. The authors demonstrate the method on two challenging systems: brookite TiO₂ nanorods, which form dense conglomerates, and CsPbBr₃ perovskite nanocrystals, which are beam‑sensitive and only ~30 nm in size.

For TiO₂, five distinct particles were tracked across a 0.20° tilt step series (DS‑1). Despite a total dose of only ~1.1 e⁻ Å⁻² per tilt, the diffraction resolution reached 3 Å⁻¹, sufficient for reliable structure solution. Four of the five particles yielded complete unit‑cell determinations with internal consistency metrics R_int < 15 % and R_meas < 18 %, far better than typical 3D ED datasets where dynamical scattering often pushes these values to 25–30 %. One particle (T‑P5) displayed two overlapping domains and was excluded, illustrating the method’s ability to detect multi‑domain contamination.

For CsPbBr₃, low‑dose conditions (≈0.15 e⁻ Å⁻² per tilt) were essential to avoid rapid ligand‑induced contamination. The SAM2 segmentation successfully isolated the weak particle signal from the noisy background, enabling extraction of 3D ED data with 70–80 % completeness. Subsequent dynamical refinement, extinction correction, and anisotropic displacement parameter modeling produced R_obs values comparable to those obtained from larger, more robust crystals.

The authors also investigated the impact of tilt‑step size by comparing a finer 0.10° series (DS‑2) with the coarser 0.20° series. The finer steps improved reflection intensity profiles and reduced R_int, confirming that dense reciprocal‑space sampling is critical for high‑quality intensity integration.

All experiments were performed with commercially available hardware: a fast event‑based direct electron detector, a standard TEM, and a desktop workstation for data processing. Acquisition times ranged from 40 to 125 minutes per tomogram, and total data‑reduction times remained within a few hours, demonstrating practical feasibility for routine laboratory use.

In summary, the presented 4D‑STEM tomography combined with automated object tracking offers three decisive advantages: (i) real‑space tracking eliminates drift and ensures that the same crystal is followed throughout the tilt series, (ii) low‑dose, fine‑step acquisition preserves beam‑sensitive specimens while delivering high‑resolution reciprocal‑space coverage, and (iii) the wide field of view permits simultaneous extraction of multiple independent 3D ED datasets from a single experiment. This approach expands the applicability of electron diffraction‑based nanocrystallography to previously inaccessible material classes, paving the way for routine structural analysis of complex, beam‑sensitive, and densely packed nanomaterials.