4DSTEM Studio

4DSTEM Studio · Reference

Bragg Disks

Applies to 4DSTEM Studio 5.4 and later

Bragg Disks finds the diffraction disks in every pattern. This page lists every control of its inspector under the name the app shows.

The detector correlates each pattern with a probe kernel and keeps the correlation peaks that pass the settings. The settings and their defaults are py4DSTEM's.

Probe kernel

Control What it does
Source Where the kernel comes from: Synthetic (a disk of the probe radius), Current CBED / ROI (the pattern on screen), File's probe (a probe stored in the file) or Vacuum scan… (a separate scan without a sample)
Probe radius The radius the synthetic kernel is built at, in detector pixels. Empty until Calibrate Origin measures it. A typed value replaces the measured one.
Measured kernel mode Flat uses the probe as it is: py4DSTEM's recommendation for bullseye and other structured probes, and it needs no radius. Sigmoid trench subtracts a disk whose edge falls smoothly to zero between the probe radius and twice it, so the kernel sums to zero, has a negative trench around the probe, and the correlation responds to the disk edge (py4DSTEM's sigmoid kernel). It is only as good as the probe radius.
Build Kernel Builds the kernel from the chosen source
Kernel The kernel in use: its source, its mode and its radius

For a ring-shaped probe the app notes where the probe's outer edge is and offers a button to use that radius.

Detector

Control What it does
Detector Classical or Neural net. The neural net proposes candidate positions on the whole pattern; the classical refinement still measures every one.
Threshold The pick threshold on the neural net's heatmap, 0.3 to 0.99. Lower accepts more candidates.
Model The neural-net model in use: Bundled, or one you fine-tuned on your own labels. Each result records which model made it.
Compare Detectors Shows a map of the peaks the two detectors do not share at each position. Available once Detect All Disks has run with each detector.

The neural net is offered when Settings › Analysis › Offer the learned (neural net) detector is on.

Disk detection

With a kernel built, the pattern on screen shows the detected disks as rings, and this section counts them:

Row How to read it
Current CBED The number of disks found in the pattern on screen
Acceptance funnel How many candidates there were and how many were accepted, then how many passed the absolute threshold, the relative threshold and the spacing rule. It shows which setting removes the disks you miss.
Peaks found After Detect All Disks: the number of disks in the whole scan
Per pattern The median and the range of disks per pattern over the scan

Detect All Disks, the toolbar's action, runs the detection over the whole scan (⌘R).

Advanced detection

Changes update the rings on the current pattern at once. Run Detect All Disks again to apply them to strain and orientation.

Setting Default What it does
Maximum peaks 70 The most disks kept for each pattern
Pattern smoothing σ 0 px Gaussian smoothing of each pattern before the correlation. Too much merges nearby features.
Correlation smoothing σ 2 px Gaussian smoothing of the correlation map before its maxima are picked. Zero turns it off.
Minimum absolute 0 The weakest correlation peak kept, on the absolute correlation scale. It depends on the data and the kernel. Zero turns it off.
Minimum relative 0.5 % The weakest peak kept, as a fraction of the reference peak. Zero turns it off.
Reference peak #1 The peak the relative threshold is measured against, by brightness rank. #1 is the brightest.
Reference outside 0 px Measures the relative threshold against the brightest candidate at least this far from the brightest one, so that a saturated direct beam does not set the scale. For a flat plateau, set at least its diameter. 0 keeps py4DSTEM's rule.
Minimum spacing 60 px The smallest distance between two disks. When two candidates are closer, the brighter one is kept. Building a kernel sets it to fit the probe unless you chose a value.
Edge exclusion 20 px Drops maxima closer than this to the edge of the detector
Correlation power 1 1 is cross-correlation (robust on noisy data), 0 is phase correlation (sharp, for patterned probes), values between mix both
Subpixel localization Fourier multicorr Pixel is fastest, parabolic is quick, Fourier multicorr refines for strain-grade positions
Fourier upsampling 16× The upsampling used by Fourier multicorr. Higher is finer and slower.

Reset Recommended Settings puts every setting back to its default.

The defaults are scaled to the detector when a dataset opens, so a small detector starts with a smaller spacing and edge exclusion than the table shows.

Training labels

This section is for fine-tuning the neural net on your own data.

Control What it does
Label centres on click While on, a click on the diffraction pane adds a disk centre at the current scan position; a click near one removes it
This position, Labelled How many centres are labelled here and in the whole scan
Clear This Position Removes every centre at the current scan position
Save to Session Writes the labels to the session file beside the dataset
Export Labels…, Import Labels… Writes the labels to a JSON file, or replaces the current labels with those of a file
Split About 30 % of the labelled positions are held out. They judge the fine-tuned model and never train it.
Train Model… Fine-tunes the bundled model on the training positions and judges it against the active model on the held-out ones. It is offered only if it is at least as good on recall and precision and better on one.

Warnings

The app says What it means
Detection failed on this pattern The detection itself failed here. The empty result is a failure, not a pattern without disks.
This pattern was truncated to the configured maximum peak count More disks passed than Maximum peaks allows
The selected reference-peak rank is absent in this pattern There are fewer peaks than the rank of Reference peak, so the relative threshold cannot be applied
Detection settings changed — rerun Detect All Disks The disks of the whole scan were made with other settings than those shown

Method

Template matching as in py4DSTEM: Savitzky et al. (2021), Microsc. Microanal. 27, 712. Hybrid correlation: Pekin et al. (2017), Ultramicroscopy 176, 170. The numbers are in the app's parity notes.

For how to judge a detection, see Troubleshooting.