Abstract
Nanoparticle assemblies can be investigated in 3 dimensions using electron tomography. However, it is not straightforward to obtain quantitative information such as the number of particles or their relative position. This becomes particularly difficult when the number of particles increases. We propose a novel approach in which prior information on the shape of the individual particles is exploited. It improves the quality of the reconstruction of these complex assemblies significantly. Moreover, this quantitative Sparse Sphere Reconstruction approach yields directly the number of particles and their position as an output of the reconstruction technique, enabling a detailed 3D analysis of assemblies with as many as 10,000 particles. The approach can also be used to reconstruct objects based on a very limited number of projections, which opens up possibilities to investigate beam sensitive assemblies where previous reconstructions with the available electron tomography techniques failed.
Introduction
Electron tomography (ET) has become the gold standard for visualizing nanomaterials in three dimensions. However, as nanoparticle assemblies grow in size and complexity—often containing thousands of individual units—standard reconstruction algorithms like the Simultaneous Iterative Reconstruction Technique (SIRT) face significant challenges.
The primary hurdle is the "missing wedge" effect. Due to physical constraints in the microscope, samples cannot be tilted through a full 180-degree range. This incomplete data leads to blurring and elongation in the 3D model, making it nearly impossible to resolve individual particles within dense clusters. Furthermore, traditional methods require manual "segmentation" to identify and label each particle for structural analysis, a task that is practically unfeasible for large-scale assemblies. Additionally, many samples are beam-sensitive; the high electron dose required for a standard tilt series (often 70+ images) can deform or destroy the assembly before data collection is complete.
Sparse Sphere Reconstruction (SSR)
To overcome these limitations, we developed the Sparse Sphere Reconstruction (SSR) technique. The core innovation lies in shifting the reconstruction objective: instead of determining the density of every 3D pixel (voxel), the algorithm solves for the precise coordinates of the particle centers.
Mathematical Framework
The SSR approach incorporates prior knowledge of the particles' shape. By assuming the assembly consists of uniform spheres of a known diameter, the 3D object ($x$) can be modeled as a convolution of a sparse map of sphere centers ($y$) and a single sphere template ($C$).
Because the number of particles is small relative to the total volume, the vector $y$ is "sparse" (containing mostly zeros). We utilize the SPGL1 solver and the ASTRA toolbox to solve for $y$ by minimizing the $L_1$ norm. Unlike standard methods that produce a blurry density map, SSR directly outputs the 3D coordinates of every particle center, bypassing the need for manual segmentation.
Experimental Validation
The SSR method was validated across three distinct datasets:
- Simulated Phantoms: A 2D hexagonal lattice was used to demonstrate robustness against reduced tilt ranges and under-sampling.
- Cobalt Iron Oxide (Co–Fe–O) Assemblies: We analyzed spherical clusters ranging from 50 nm to 300 nm, containing between 70 and 9,301 nanoparticles (~9 nm each).
- Gold (Au) Nanoparticles in Polystyrene: This beam-sensitive system demonstrated that SSR can produce accurate 3D models from as few as 8 projections, compared to the standard 75 required by SIRT.
Key Insights and Structural Analysis
By using shape as a mathematical constraint, SSR effectively "fills in" the data lost to the missing wedge. The algorithm places mathematically perfect spheres where the projection data provides the strongest support, eliminating artifacts typical of traditional voxel-based reconstructions.
This quantitative precision allowed for a detailed study of assembly symmetry. Our analysis revealed that as clusters grow, their internal symmetry shifts. While small clusters exhibit icosahedral symmetry, larger assemblies (9,000+ particles) adopt a rhombicosidodecahedral outer morphology with a distorted "Mackay icosahedron" core. These findings confirm theoretical models of hard-sphere self-assembly under spherical confinement and allow for the identification of internal defects, such as twinning planes and hcp stacking within fcc-dominated clusters.
Conclusion and Future Directions
The SSR technique enables the 3D mapping of assemblies with up to 10,000 particles with unprecedented accuracy and efficiency. By reducing the required number of projections by nearly an order of magnitude, it opens new avenues for studying beam-sensitive soft matter and biological-nanoparticle hybrids.
Currently, SSR is optimized for monodisperse spherical particles. Future work will focus on:
- Polydispersity: Adapting the algorithm to handle spheres of multiple diameters.
- Dictionary Learning: Extending the method to non-spherical shapes (e.g., rods or cubes) using a pre-defined dictionary of templates, though this may require a higher number of projections to resolve orientation.
Summary updated