Non-local Denoising for 4D STEM Orientation Mapping
David J. Rowenhorst, Patrick G. Callahan, Yichen Yang, Justin Nakamura, William C. Lenthe, Josh Kacher · Microscopy and Microanalysis · 2025
The advent and broad application of 4D-STEM in electron microscopy has enabled a wide range of data-intensive materials analysis, including strain mapping and structure quantification. Orientation mapping has also been demonstrated using 4D-STEM as it provides higher spatial resolution than traditional electron backscatter diffraction (EBSD)-based approaches. However, dynamical diffraction effects, which are present in most TEM samples, lead to internal diffraction spot structuring that reduce the accuracy and reliability of orientation mapping. In addition, off-zone 4D-STEM patterns are typically dominated by only a few bright diffraction spots, with many of the spots that provide critical information for indexing having intensities that are very near the noise levels of the detector, adding further difficulty to orientation mapping. Precession electron diffraction can address this issue, but at the cost of additional instrumentation needs, slower pattern collection rates, and decreased spatial resolution. Alternatively, frame averaging can reduce the noise, but at the cost of making long experiments even more time consuming. In this work, we adapt the NLPAR denoising algorithm that was previously developed for increasing the quality of EBSD patterns before indexing [1]. The algorithm leverages the fact that much of the data within a 4D STEM map is highly repetitive, and thus there is a high likelihood finding similar patterns within the scan data to the pattern of interest. Here, the L2 norm of the difference between the patterns is evaluated for all patterns within a user defined search window and is used as a measure of pattern similarity between the search window patterns and the pattern of interest (with high similarity showing a low L2 value). The processed pattern is a weighted average of all the patterns within the window, with the weights determined by an exponential decay of the similarity metric. For 4D-STEM, there are some key considerations that must be applied for this to work properly. Because the transmitted beam is present in all patterns, and generally saturates the detector, it shows very little variation between patterns, and thus leads to an artificially high signal similarity. To prevent this artificial similarity, the transmitted beam is masked out of the patterns and thus omitted from the NLPAR calculation. In a similar fashion, any high intensity spots that saturate the detector will not display the typical detection variability, and will again provide artificially high pattern similarities, and thus are dynamically masked out from the calculation point by point. Figure 1 below shows a comparison of STEM patterns collected from an ultrafine-grained Au thin film before and after processing with the NLPAR algorithm. The faint spots that were nearly undetectable within the detector noise are now clearly visible within the diffraction pattern. The example intensity profile shows that the noise between diffraction peaks is significantly reduced, and that there is relatively little change to the position and intensities of the peaks. Figure 2 shows the final orientation map after indexing, wherein individual diffraction patterns are automatically indexed by comparing to a dictionary of kinematically simulated templates using an approach similar to [2] and [3] [4]. Original and NLPAR processes diffraction patterns (intensities are logarithmically scaled for display), and example intensity profiles for the original and processed patterns (linear intensity in arbitrary units). Orientation maps after diffraction pattern indexing. The reliability of the indexing is significantly improved after NLPAR processing of the diffraction patterns.