The precise three-dimensional (3D) chemical mapping of nanoscale materials is crucial for understanding their properties and designing new materials. However, achieving this at high resolution has been a significant hurdle. Traditional methods relying on inelastic scattering for chemical information require high electron doses, often exceeding the radiation tolerance of the sample, thus limiting both resolution and the range of materials that can be studied. This necessitates a new approach to 3D chemical imaging. This research aims to overcome these limitations by introducing fused multi-modal electron tomography, a technique which combines information from elastic and inelastic scattering to significantly improve the signal-to-noise ratio (SNR) and resolution in 3D chemical mapping. The technique's potential impact extends to diverse fields of materials science, allowing researchers to investigate the complex relationship between structure and chemistry in a broader range of materials.

Previous attempts at 3D chemical tomography using electron energy loss spectroscopy (EELS) or energy-dispersive X-ray spectroscopy (EDX) have faced significant challenges due to the low signal-to-noise ratio (SNR) of inelastic scattering events. These techniques often require high electron doses, which can damage the sample and limit resolution. While progress has been made in 3D structural imaging using elastic scattering signals, combining these elastic structural data with inelastic chemical information has been a major challenge. The introduction of scanning transmission electron microscopy (STEM) tomography represented a milestone in 3D imaging, but the dose constraints remained a major limitation. The authors cite previous works that attempted 2D multi-modal data fusion, showing promise in reducing dose requirements but without extension to the crucial 3rd dimension. The current study aims to address the need for high-resolution 3D chemical tomography through a novel fusion of multimodal data in the 3D space, overcoming the limitations of previous approaches.

The researchers developed fused multi-modal electron tomography, a technique that combines information from both elastic (high-angle annular dark-field, HAADF) and inelastic (EELS/EDX) scattering. HAADF imaging provides high-resolution structural information, while EELS/EDX offers chemical specificity. The key innovation lies in fusing these two modalities during the 3D reconstruction process. This fusion is achieved by solving an optimization problem that considers both the HAADF and chemical signals, while also incorporating sparsity regularization to reduce noise and enhance resolution. The optimization problem is formulated as: minₓ,b ½||∑ᵢ ZᵢAxᵢ − bᵢ||² + λ₁∑ⱼ ||Aⱼx − bⱼ||² + λ₂||x||tv where xᵢ represents the reconstructed 3D chemical distributions, bᵢ are the measured HAADF intensities, Aᵢ and Aⱼ are forward projection operators, Z, λ₁, and λ₂ are regularization parameters, and ||x||tv represents total variation regularization. This framework allows for the acquisition of significantly fewer chemical projections while utilizing the rich structural information provided by numerous HAADF projections. The method is tested on multiple materials including Au-Fe₂O₃ superlattice nanoparticles, Co₃O₄-MnO₂ core-shell nanocrystals, ZnS-Cu₂O₄ nanoparticles, and Cu-SiC nanoparticles, as well as simulated datasets. Specimen preparation and tilt-series acquisition are carefully described, detailing the microscope settings and parameters used. The alignment of HAADF and EELS/EDX tilt series is performed using iterative image registration techniques, followed by background subtraction and projection matching for fine alignment. The reconstruction process is detailed, explaining the optimization algorithms, regularization parameters, and convergence criteria. Bayesian optimization is employed to efficiently determine the optimal regularization parameters for the different materials and simulated scenarios. The resolution of the reconstructed chemical maps is assessed in both real and reciprocal space.

The study demonstrates the ability to achieve sub-nanometer (1nm) 3D resolution in chemical mapping across a variety of material systems, using significantly lower electron doses compared to conventional methods. This is made possible by the synergistic combination of HAADF and EELS/EDX data. Specific findings include: * **High-Resolution 3D Chemical Mapping:** The fused multi-modal approach successfully resolves the 3D chemical distributions in Au-Fe₂O₃ superlattice nanoparticles with a resolution of 0.8 nm × 0.8 nm × 1.1 nm. The ability to distinguish Fe nanoparticles (1.0 ± 1.1 nm) from Au nanoparticles (3.9 ± 0.4 nm) in the superlattice structure is clearly demonstrated. * **Core-Shell Nanocrystal Analysis:** In Co₃O₄-MnO₂ core-shell nanocrystals, the technique reveals intricate details of the core-shell interface, including the intrusion of MnO₂ strands into the Co₃O₄ core, providing insights into the growth mechanism. * **Dose Reduction:** The fusion of HAADF and chemical data results in a 100-fold reduction in electron dose compared to traditional chemical tomography, making it applicable to a broader range of radiation-sensitive materials. * **Improved SNR:** The fused multi-modal approach substantially improves the SNR of the chemical maps, leading to better visualization and quantification of chemical composition. * **Stoichiometry Determination:** The method allows for accurate 3D stoichiometry determination without prior knowledge of elastic scattering cross-sections. Simulations show that the stoichiometric precision of the multi-modal method is four times better than traditional methods. * **Validation through Simulations:** Simulations using synthetic nanocrystals validate the method's efficacy, demonstrating significant improvements in reconstruction accuracy and resolution compared to conventional chemical tomography. * **Broad Applicability:** The technique is shown to be effective for several materials including Au-Fe₂O₃ nanoparticles, Co₃O₄-MnO₂ nanocrystals, ZnS-Cu₂O₄ heterostructures, and Cu-SiC nanoparticles, thus demonstrating broad applicability across materials classes.

The results demonstrate a significant advancement in the field of 3D chemical imaging. The fused multi-modal electron tomography technique presented overcomes the limitations of traditional methods by leveraging the complementary strengths of HAADF and EELS/EDX imaging. The substantial reduction in electron dose achieved is particularly important, expanding the range of materials amenable to high-resolution 3D chemical analysis. This technique has broad implications across materials science, enabling deeper understanding of complex nanomaterials and providing crucial insights into structure-property relationships. The ability to quantify 3D stoichiometry adds another dimension to the characterization toolkit. The detailed analysis of core-shell nanocrystals exemplifies the capability of this approach to reveal fine details of interfaces and growth mechanisms, which is not accessible through conventional methods. The validation using simulations further confirms the robustness and accuracy of the technique. The demonstrated improvement in both resolution and SNR compared to traditional methods establishes its superiority for 3D chemical imaging.

This research successfully demonstrates a novel fused multi-modal electron tomography method capable of achieving sub-nanometer 3D resolution in chemical mapping with significantly reduced electron dose. The technique's success is validated across diverse materials, highlighting its potential for widespread application in materials science. Future research could focus on further algorithmic optimization and exploration of different inelastic scattering techniques to expand its capabilities. Investigating the application of this technique to even more challenging materials and exploring possibilities of automation would also be valuable contributions.

The current study primarily focuses on relatively well-defined nanostructures. The applicability of this technique to more amorphous or disordered materials needs further investigation. The computational cost of the reconstruction process, while mitigated by Bayesian optimization, could still be a limiting factor for extremely large datasets. The method's performance might also be affected by the choice of regularization parameters and their optimization; although Bayesian Optimization is employed, the selection of acquisition function and kernel for GP would influence the results.