NLSTEM: Non-local denoising for enhanced 4D-STEM pattern indexing

This paper introduces NLSTEM, a non-local denoising post-processing technique that significantly improves 4D-STEM pattern indexing rates by averaging similar diffraction patterns to enhance signal-to-noise ratios, particularly benefiting ion-damaged samples with curved lattices without requiring beam precession.

The Gist

The Big Picture: Seeing the Invisible Crystal World

Imagine you are trying to figure out the layout of a massive, complex city made entirely of tiny, invisible Lego bricks. You have a super-powerful microscope (a 4D-STEM) that takes a picture of the "shadow" (diffraction pattern) cast by the bricks at every single spot in the city.

The goal is to map out the city: Where are the streets? Where are the buildings? Which way are the bricks facing? This is called orientation mapping.

However, there's a problem. The city is so crowded and the bricks are so small that the shadows are often blurry, messy, and full of static (noise). It's like trying to read a street sign through a foggy, dirty window. Sometimes, the computer software can't tell what the sign says, so it just gives up and leaves a blank spot on the map.

The Old Way: The "Precession" Problem

To fix this blurry window, scientists used to shake the camera (a technique called beam precession). Imagine shaking a camera while taking a photo to blur out the dust motes. It works, but it's expensive, requires special hardware, and makes the photo slightly less sharp (lower resolution). It's like buying a very expensive, heavy camera lens just to get a clearer picture.

The New Solution: NLSTEM (The "Smart Crowd-Sourcer")

This paper introduces a new software trick called NLSTEM. Instead of buying new hardware, they invented a smarter way to process the photos after they are taken.

Think of it like this:
Imagine you are trying to identify a specific face in a crowd, but the person is wearing a mask and the lighting is bad.

• The Old Way: You squint harder at that one person.
• The NLSTEM Way: You look at that person, and then you look at everyone standing within a 10-foot radius. You ask yourself, "Who looks most like this person?" You find 80 people who look very similar. Instead of just looking at the one blurry face, you take a "mental average" of all 80 faces.

Because the noise (the static) is random, it cancels out when you average 80 people together. But the actual face (the crystal structure) stays the same. Suddenly, the face becomes crystal clear.

How It Works (The Magic Sauce)

The algorithm is "Non-Local." This is the key difference.

• Local Averaging (The Old Way): If you are standing on a street corner, you only look at the 4 people immediately next to you. If one of them is wearing a red hat (a different crystal orientation), your average gets messed up. You lose the detail of the street corner.
• Non-Local Averaging (NLSTEM): The algorithm looks at the entire city map. It finds people who look like the person you are studying, even if they are on the other side of town. It ignores the people who look different (the red hats) and only averages the people who look the same.

This allows the software to clean up the noise without blurring the edges of the buildings. You get a clear picture of the tiny details and the big picture.

The Surprising Discovery: Broken is Better?

The researchers tested this on gold and nickel films. They did something weird: they shot the gold films with high-energy ions (basically, they damaged the crystal structure with radiation).

Usually, damaging a crystal makes it harder to read. But here, the damaged gold actually got easier to map after using NLSTEM.

Why?
Imagine the gold crystals are like a flat sheet of paper. When you damage it, it starts to curl and warp slightly.

• In the old method, a warped sheet looks like a mess.
• In the NLSTEM method, the algorithm averages the "warped" neighbors. It turns out that this warping acts like a natural "shaking" of the camera (similar to the expensive hardware method mentioned earlier). It smooths out the weird physics of the light, making the pattern easier to read.

It's like finding out that a slightly crumpled piece of paper is actually easier to read than a perfectly flat one when you use this specific new reading glasses.

Why This Matters

1. Cheaper: You don't need to buy expensive hardware upgrades. You just need a computer.
2. Sharper: You can see tiny details (like thin walls between grains) that used to get blurred out by older software.
3. Faster: It works on data you already have. You can take a "bad" scan and turn it into a "good" map later.

The Bottom Line

The authors created a "smart filter" for electron microscope images. It finds similar patterns in a dataset, averages them to remove the static noise, and reveals the hidden crystal structure. It's like turning a grainy, black-and-white security camera feed into a high-definition, color movie, allowing scientists to see the microscopic world with incredible clarity.

Technical Summary

1. Problem Statement

4D-STEM (Four-Dimensional Scanning Transmission Electron Microscopy) has revolutionized microstructure quantification by enabling rapid orientation and phase mapping. However, accurate pattern indexing (determining crystal orientation and phase) faces significant challenges:

• Low Signal-to-Noise Ratio (SNR): In nanocrystalline materials or heavily damaged samples, diffraction patterns are often noisy, leading to low indexing success rates.
• Pattern Mixing: In thin films with grain sizes comparable to the probe size, a single probe position may intersect multiple grains, creating mixed diffraction patterns that standard template matching algorithms struggle to resolve.
• Limitations of Current Solutions:
  • Hardware: Beam precession (PED) improves indexing by averaging dynamical effects but increases cost, acquisition time, and often reduces spatial resolution.
  • Software: Standard template matching lacks dynamical diffraction information. Existing denoising methods like Nearest Neighbor Pattern Averaging (NPAR) smooth data indiscriminately, often blurring sharp features and small-scale microstructural details (e.g., thin twin lamellae).

2. Methodology: The NLSTEM Algorithm

The authors introduce NLSTEM (Non-Local denoising for 4D-STEM), a post-processing algorithm adapted from the Non-Local Pattern Averaging and Reindexing (NLPAR) method used in Electron Backscatter Diffraction (EBSD).

Core Logic:
Instead of averaging only immediate neighbors, NLSTEM calculates the similarity between a target diffraction pattern ($p_i$) and all other patterns within a defined search window ($W$). It then performs a weighted average based on this similarity.

Key Technical Adaptations for 4D-STEM:

1. Noise Modeling: Unlike EBSD (dominated by Gaussian readout noise), 4D-STEM patterns (especially from direct detection cameras) are dominated by Poisson noise, where noise magnitude scales with the square root of intensity.
   • Solution: The algorithm scales pattern intensities by their square root ($q = \sqrt{p}$) before calculating distances to standardize noise across intensity levels, then scales the result back ($p' = (q')^2$).
2. Handling Saturation: 4D-STEM patterns often contain saturated pixels (e.g., the direct beam). Comparing these creates artificial similarity.
   • Solution: A mask excludes the direct beam. Furthermore, the distance metric excludes any pixel where either pattern is within 1% of the maximum intensity in the window.
3. Weighting Function: The algorithm uses an L2 norm distance metric to calculate a normalized weighting factor ($w$). Patterns with low distance (high similarity) receive higher weights. The decay of weights is controlled by a parameter $\lambda$ (optimized to 1.07 in this study).
4. Search Window: A square sliding window with a radius of 4 (80 neighbors) is used to find similar patterns, balancing computational cost with the likelihood of finding true structural matches.

3. Key Contributions

• Algorithm Development: Adaptation of the NLPAR algorithm specifically for the high dynamic range and Poisson noise characteristics of 4D-STEM data.
• Preservation of Features: Unlike simple nearest-neighbor averaging, NLSTEM preserves sharp edges and small-scale features (like thin twins) because it only averages patterns that are statistically similar, avoiding the blurring of distinct microstructural boundaries.
• Open Source Implementation: The authors provide the source code (compatible with py4DSTEM and Gatan STEMx), making the technique accessible to the broader community.

4. Results

The algorithm was validated on two material systems: Nanocrystalline Nickel (Ni) and Ultrafine-grained Gold (Au) subjected to ion irradiation.

• Nanocrystalline Ni:
  • Before NLSTEM: Indexing rate was 90.6%, with failures concentrated at grain boundaries and triple junctions due to pattern mixing.
  • After NLSTEM: Indexing rate increased to >99.9%. The resulting maps showed contiguous grain interiors with sharp boundaries, and the signal-to-noise ratio in diffraction patterns improved significantly, revealing faint Bragg discs.
• Irradiated Au (0, 1, and 5 dpa):
  • Observation: As ion damage increased (up to 5 displacements per atom), the raw indexing rate dropped significantly due to lattice strain and defect-induced diffuse scattering.
  • NLSTEM Effect: Post-processing recovered indexing coverage at all damage levels.
  • Surprising Finding: The indexing rate for the 5 dpa (heavily damaged) sample after NLSTEM processing was higher than that of the pristine (0 dpa) sample.
  • Explanation: The authors suggest an "inverted PED" effect. Ion irradiation induces local lattice curvature and rotations. NLSTEM averaging over these curved lattices effectively integrates over small angular ranges, damping dynamical diffraction oscillations and producing quasi-kinematic patterns that are easier to index.

5. Significance and Comparison

• Superiority over NPAR: Comparisons on annealed Ni films showed that while NPAR failed to resolve thin twin lamellae (especially at coarser step sizes), NLSTEM successfully resolved features as small as the scan step size (10 nm) while maintaining continuity.
• Impact on Microscopy:
  • Enables high-fidelity orientation mapping in materials where it was previously impossible (e.g., nanocrystalline or heavily irradiated materials).
  • Reduces the need for hardware modifications like beam precession, saving time and cost while maintaining spatial resolution.
  • Improves the reliability of template matching by enhancing the visibility of high-order reflections and faint peaks that were previously lost in noise.

Conclusion:
NLSTEM represents a significant advancement in 4D-STEM data processing. By leveraging non-local denoising tailored to the specific noise characteristics of electron diffraction, it dramatically improves indexing rates and preserves critical microstructural details, offering a powerful software-based solution for complex materials characterization.