Unsupervised anomaly detection in MeV ultrafast electron diffraction

This paper presents a fully unsupervised methodology using a convolutional autoencoder to detect and filter anomalous MeV ultrafast electron diffraction patterns caused by beam instabilities, achieving high accuracy with minimal training data and rapid processing times.

The Gist

The "Smart Sorter" for Blurry Photos: A Simple Explanation

Imagine you are trying to take the perfect group photo of a thousand people, but every time you snap the picture, someone sneezes, a camera shakes, or a flash goes off at the wrong time. If you just stack all those photos on top of each other to make one "super photo," the sneezes and shakes will blur the final image, making it hard to see the details.

This is exactly the problem scientists face with MeV Ultrafast Electron Diffraction (MUED).

The Problem: The "Sneezing" Electron Beam

Scientists use MUED to take incredibly fast "snapshots" of how atoms move inside materials. They fire a beam of electrons at a sample to see its structure. To get a clear picture, they take thousands of these snapshots and average them together.

However, the electron beam is a bit unstable. Sometimes it wobbles, drifts, or gets jittery. When this happens, the resulting "snapshot" (called a diffraction pattern) looks weird or distorted. If you include these "bad" snapshots in your final average, the whole picture gets blurry, and you miss the tiny, important details of how the material is changing.

Traditionally, scientists had to look at thousands of these images one by one to find the bad ones. It's like trying to find a few spoiled apples in a barrel of 1,000 by looking at each one with a magnifying glass. It's slow, boring, and prone to human error.

The Solution: The "Auto-Recall" Robot

The authors of this paper built a fully unsupervised AI robot (a Convolutional Autoencoder) to do the sorting for them. Here is how it works, using a simple analogy:

1. The Training Phase: Learning the "Normal"
Imagine you show the robot 100 photos of a perfect, clear crystal. You don't tell it "this is good" or "this is bad." You just let it look at them.

• The robot tries to memorize what a perfect crystal looks like.
• It then tries to redraw the crystal from memory.
• If the drawing looks exactly like the original, the robot says, "I got it right!"
• If the drawing is blurry or missing parts, the robot knows, "I'm struggling to remember this."

2. The Testing Phase: Spotting the Weirdos
Now, the robot is shown 1,500 new photos, some of which are perfect and some of which are distorted by the "sneezing" electron beam.

• The Good Photos: The robot easily recognizes them and redraws them perfectly. The "error" (the difference between the original and the redraw) is tiny.
• The Bad Photos: The robot has never seen these weird distortions before. It tries to redraw them, but it fails miserably because they don't match its memory of a "normal" crystal. The "error" is huge.

3. The Decision: The Probability Score
Instead of just saying "Good" or "Bad," the robot gives a confidence score.

• "This image is 99% likely to be normal." -> Keep it.
• "This image is 99% likely to be faulty." -> Throw it away.
• "This image is 50/50." -> Flag it for a human to double-check.

This is the "unsupervised" magic: The robot figured out what "normal" looks like on its own, without anyone ever telling it what a "bad" photo looked like.

Why This is a Big Deal

• Speed: The robot can process an image in about 1 second. A human would take minutes.
• Accuracy: It catches the bad images with a false alarm rate of less than 0.4%. That means out of 1,000 good photos, it will rarely throw one away by mistake.
• Efficiency: It was trained on just 100 images. Usually, AI needs thousands or millions to learn. This proves the method is very efficient.
• Versatility: This isn't just for electron beams. Any time a machine takes thousands of pictures and some get messed up by glitches (like a shaky camera or a dirty lens), this "Smart Sorter" can help clean up the data.

The Bottom Line

The scientists built a digital bouncer for their data. Instead of humans manually checking thousands of blurry photos to find the glitches, this AI learns what a "perfect" photo looks like and instantly kicks out the ones that don't fit the party. This allows scientists to get sharper, clearer pictures of the atomic world, helping them understand how materials behave in the blink of an eye.

Technical Summary

1. Problem Statement

MeV Ultrafast Electron Diffraction (MUED) is a pump-probe technique used to study the dynamic structural evolution of materials with high temporal and spatial resolution. However, the technique faces significant challenges:

• Signal-to-Noise Ratio (SNR): To achieve sufficient SNR, diffraction patterns must be averaged over thousands of "shots" (individual electron beam pulses).
• Beam Instabilities: Shot-to-shot instabilities in the relativistic electron beam cause distortions in individual diffraction patterns.
• Consequence: When these distorted (anomalous) patterns are averaged into the dataset, they degrade the resolution of the final image, potentially obscuring subtle structural changes in the material.
• Current Limitation: Manually inspecting thousands of images to remove faulty ones is time-consuming and impractical. Existing machine learning approaches often require labeled data, which is difficult to obtain for this specific application.

The core problem is the need for a fully unsupervised, automated method to detect and filter out anomalous diffraction patterns from large datasets without prior manual labeling.

2. Methodology

The authors propose a pipeline combining image preprocessing, Convolutional Autoencoders (CAE), and probabilistic statistical analysis.

A. Data Preprocessing

• Tiling: Raw 512x512 pixel diffraction images are normalized and divided into overlapping 80x80 pixel tiles.
• Filtering: Tiles containing only background (white noise) are filtered out during training to ensure the model learns the features of Bragg peaks. This is achieved by counting connected pixels above a threshold (derived from the tile's median amplitude).
• Rationale: This reduces computational load and focuses the model on the signal-rich regions of the diffraction pattern.

B. Convolutional Autoencoder (CAE)

• Architecture: A lightweight CAE is used. It consists of an encoder (3 convolutional layers with MaxPooling) and a decoder (3 transpose convolutional layers with upsampling).
  • Input: 80x80 pixel tile.
  • Bottleneck: A latent space of dimension 256.
  • Output: Reconstructed 80x80 tile.
• Training: The CAE is trained only on normal (fault-free) images (100 images). The objective is to minimize the Mean Squared Error (MSE) between the input and the reconstructed output using an Adam optimizer.
• Mechanism: The model learns to reconstruct "normal" diffraction patterns effectively. Because anomalies (caused by beam instabilities) follow a different distribution, the CAE fails to reconstruct them accurately, resulting in high reconstruction errors.

C. Anomaly Detection & Probabilistic Modeling

• Residual Analysis: The detection metric is the MSE of the reconstruction error across all pixels of a tile.
• Statistical Modeling: The distribution of reconstruction errors is modeled as a mixture of two probability density functions:
  • $p(e|N)$: Likelihood of error given the image is Normal.
  • $p(e|A)$: Likelihood of error given the image is an Anomaly.
• Distribution Choice: The error distributions are modeled using Rice distributions (or biased Rayleigh for anomalies).
• Parameter Estimation: The parameters (shape, scale, bias) and the prior probabilities ($p(N)$ and $p(A)$) are estimated using the L-BFGS-B algorithm to minimize the negative log-likelihood of the observed errors.
• Output: The system calculates the posterior probability $p(N|e)$ that a specific image is normal.
  • Images with $p(N|e) \approx 1$ are accepted.
  • Images with $p(N|e) \approx 0$ are rejected.
  • Images near 0.5 are flagged for manual inspection.

3. Key Contributions

1. Fully Unsupervised Framework: The method requires no labeled data. The system learns the definition of "normal" solely from the data itself, eliminating the need for tedious manual labeling of training sets.
2. Probabilistic Confidence: Unlike binary classifiers, this method provides a posterior probability for each image. This allows users to make informed decisions, specifically flagging ambiguous cases (probability $\approx 0.5$) for human review while automatically discarding clear outliers.
3. Efficiency: The method is computationally lightweight. It can be trained on a small dataset (100 images) and processes images rapidly.
4. Generalizability: While tuned for MUED, the framework is applicable to other diffraction techniques (e.g., in-situ transmission electron diffraction) where instrumental instabilities introduce faulty images into large datasets.

4. Experimental Results

• Dataset:
  • Training: 100 normal images (5,492 tiles).
  • Testing: 1,521 images (including 615 faulty images, ~40% anomaly rate in the initial test set).
• Performance Metrics:
  • False Positive Rate (FPR): Between 0.2% and 0.4% (specifically $2.2 \times 10^{-3}$ to $4.4 \times 10^{-3}$ depending on the dataset composition).
  • Detection Rate: 100% (all faulty images were identified).
  • Speed:
    • Training time: ~10 seconds per image (total < 2 minutes for the set).
    • Test time: ~1 second per image.
• Robustness: The method was tested with a reduced anomaly rate (2% faulty images) and maintained high performance, with optimized parameters aligning closely with the actual data distribution.
• Limitations: The method struggles with subtle, high-amplitude artifacts that deviate significantly from the learned distribution but are not "noisy" in the traditional sense (e.g., specific detector defects).

5. Significance

This work represents a significant step forward in the automation of ultrafast electron diffraction experiments. By enabling the real-time or near-real-time removal of faulty data, the methodology directly improves the signal-to-noise ratio of averaged diffraction patterns. This leads to:

• Higher Resolution: Clearer structural evolution data.
• Increased Throughput: Researchers can process large datasets without manual intervention.
• System Diagnostics: The detection rate can serve as a metric for the health and stability of the electron beam source.

The approach effectively bridges the gap between deep learning capabilities and the practical constraints of experimental physics, where labeled data is scarce and data volume is massive.