Unsupervised Machine-Learning Pipeline for Data-Driven Defect Detection and Characterisation: Application to Displacement Cascades

This paper presents a fully unsupervised machine learning pipeline that integrates SOAP descriptors, autoencoders, UMAP, and HDBSCAN to automatically detect, classify, and characterize defect morphologies in displacement cascades across various materials without requiring manual template or threshold tuning.

The Gist

The Big Picture: Finding "Bad Apples" in a Giant Barrel

Imagine you have a massive, perfect barrel of apples (representing a solid metal material). Suddenly, a cannonball (a high-energy neutron) smashes into the barrel. This causes a chaotic chain reaction: apples bump into each other, some get squashed, some fly out of their spots, and some get crushed into weird shapes.

In the world of materials science, this is called a displacement cascade. It happens in a fraction of a second (picoseconds) and leaves behind "defects"—atoms that are no longer in their perfect, orderly spots. These defects are the "bad apples" that eventually make the metal brittle or weak.

The problem? The barrel has billions of apples. Finding the few damaged ones by looking at them one by one is impossible for a human. Traditional computer methods try to find these bad apples by looking for specific shapes (like "is this apple round?"). But if the apple is slightly squashed or weirdly shaped, those old methods often miss it or get confused.

This paper introduces a new, super-smart detective system that doesn't need to know what a "bad apple" looks like beforehand. It just knows what a "perfect apple" looks like, and it flags anything that doesn't fit that pattern.

---

The Detective Team: How the Machine Learning Pipeline Works

The authors built a four-step workflow to find and sort these defects. Think of it as a high-tech assembly line for sorting apples.

Step 1: The "Fingerprint Scanner" (SOAP Descriptors)

First, the computer takes a snapshot of every single atom in the metal. Instead of just looking at the atom, it creates a unique fingerprint for the neighborhood around that atom.

• The Analogy: Imagine taking a photo of a person and their 12 closest friends. The "fingerprint" describes exactly how those friends are standing, how close they are, and the angles between them. Even if the person moves slightly, the fingerprint captures the exact vibe of the crowd.
• The Tool: They use a method called SOAP to turn these complex 3D neighborhoods into a simple list of numbers (a vector).

Step 2: The "Memory Test" (Autoencoder Neural Network)

Next, they train a computer brain (a neural network) on a picture of a perfect, undamaged metal. The computer learns to look at a fingerprint and say, "I know this! This is a perfect neighborhood."

• The Analogy: Imagine a security guard who has memorized the face of every employee in a building. When someone walks in, the guard tries to match their face to the memory.
• The Trick: The computer is an Autoencoder. It tries to "reconstruct" the fingerprint it sees. If it sees a perfect atom, it can rebuild the fingerprint perfectly. But if it sees a damaged atom (a weird crowd of friends), it struggles to rebuild it. The "error" in the reconstruction is huge.
• The Result: The computer flags these high-error atoms as Outliers (the bad apples). It does this without ever being told what a "defect" looks like; it just knows what "normal" looks like.

Step 3: The "Map Maker" (UMAP)

Now, the computer has a list of thousands of "bad apples." But they are all different! Some are just one atom out of place; others are huge clusters of crushed atoms. The computer needs to sort them.

• The Analogy: Imagine you have a giant pile of mixed-up toys. You want to group them, but they are in a 10-dimensional box that you can't see. UMAP is like a magical map maker that flattens that 10D box onto a 2D piece of paper.
• The Magic: It places similar toys next to each other. Atoms with similar "badness" end up in the same neighborhood on the map.

Step 4: The "Group Organizer" (HDBSCAN)

Finally, the computer looks at the map and draws circles around the clusters.

• The Analogy: A party host walks into the flattened map and says, "You guys are all standing together, so you must be a group!"
• The Result: The computer automatically groups the outliers into distinct categories:
  • Group A: Atoms missing from their spot (Vacancies).
  • Group B: Atoms squeezed into extra spots (Interstitials).
  • Group C: Huge, messy piles of atoms (Complex clusters).
  • Group D: Weird, specific shapes (like icosahedrons, which are like 20-sided dice made of atoms).

---

What Did They Find?

The team tested this on three different metals: Nickel (Ni), a steel alloy (FeNiCr), and Zirconium (Zr).

1. It Works Everywhere: The system successfully found the "bad apples" in all three metals, even though they have different atomic structures (some are cubic, some are hexagonal).
2. It's a Better Detective: They compared their new AI system against old-school methods (like checking if atoms are in a perfect grid).
   • Old methods were like looking for a specific shape. If the shape was slightly off, they missed it.
   • The AI caught almost everything the old methods found, plus it found more subtle defects that the old methods missed.
3. It Sorts by Size and Type: The AI didn't just find the defects; it sorted them. It could tell the difference between a single missing atom and a massive cluster of 500 atoms, and it could tell if a cluster was made of "missing" atoms or "extra" atoms.
4. The "Icosahedral" Surprise: In Nickel, they found complex structures that look like 20-sided dice (icosahedrons). The old methods (PTM) could only spot the very center of these dice. The AI spotted the center and the distorted shell of atoms surrounding it, giving a much clearer picture of the damage.

Why Does This Matter?

Imagine you are a nuclear engineer. You need to know exactly how much damage radiation does to a reactor wall so you can predict when it will break.

• Before: You had to use a blunt instrument to guess the damage, often missing the subtle cracks that start the big breaks.
• Now: You have a laser-guided scanner that maps every single crack, sorts them by type, and tells you exactly how big they are.

This new workflow is unsupervised, meaning it doesn't need a human to teach it what to look for. It learns the rules of "normal" and automatically finds the "abnormal." This makes it a powerful tool for designing safer, longer-lasting materials for nuclear reactors and space exploration.

In a Nutshell

The authors created a smart, self-teaching system that scans a metal, learns what "perfect" looks like, and then automatically finds, counts, and sorts every single flaw caused by radiation, all without needing a human to tell it what a flaw looks like. It's like having a detective that can read the mind of the metal itself.

Technical Summary

1. Problem Statement

Radiation-induced damage in structural materials (e.g., for nuclear reactors) begins with displacement cascades, where energetic particles (Primary Knock-on Atoms, PKAs) collide with lattice atoms, generating a complex mix of point defects (vacancies, interstitials) and extended defects (loops, clusters).

• The Challenge: Characterizing these "primary damage" states is difficult experimentally due to femto- to picosecond timescales. While Molecular Dynamics (MD) simulations can model these events, identifying and classifying the resulting defects is non-trivial.
• Limitations of Conventional Methods: Traditional defect identification tools (Centrosymmetry Parameter, Common Neighbor Analysis, Polyhedral Template Matching, Dislocation Extraction Algorithm, etc.) rely on specific topological patterns or perfect lattice references. They often struggle with:
  • Subtle distortions caused by thermal fluctuations or strain.
  • Complex or unknown defect morphologies not present in their structural databases.
  • High computational costs and scalability issues.
  • The need for manual threshold tuning.

2. Methodology

The authors propose a fully unsupervised machine learning (ML) workflow that detects and classifies defects directly from MD data without prior knowledge of defect types. The pipeline consists of four sequential stages:

A. Data Encoding (SOAP Descriptors)

• Input: Local Atomic Environments (LAE) from MD simulations of displacement cascades in Ni, FeNiCr (austenitic stainless steel), and Zr (hcp).
• Descriptor: The Smooth Overlap of Atomic Positions (SOAP) vector is used to encode the local environment of every atom. These descriptors are invariant to rotation, translation, and atomic indexing.
  • Parameters: Cutoff radii of 4.7 Å (fcc) and 5.8 Å (hcp) with specific radial and spherical harmonic basis functions.

B. Anomaly Detection (Autoencoders)

• Training: An Autoencoder (AE) neural network is trained only on SOAP vectors from defect-free reference structures (equilibrated at 100 K).
• Mechanism: The AE learns to compress and reconstruct normal atomic environments.
• Outlier Identification: When applied to cascade snapshots, atoms with high reconstruction errors (Mean Squared Error, MSE) are flagged as outliers. These outliers represent atoms in perturbed environments (defects).
• Thresholding: A threshold is selected based on the log-log distribution of reconstruction errors, separating the bulk of low-error atoms from the heavy-tailed regime of defects.

C. Dimensionality Reduction (UMAP)

• The SOAP vectors of the identified outlier atoms are projected into a low-dimensional latent space using Uniform Manifold Approximation and Projection (UMAP).
• Unlike PCA, UMAP preserves both local neighborhood relationships and global data structure, ensuring that distinct defect types remain separable in the 2D/3D embedding.

D. Clustering (HDBSCAN)

• Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) is applied to the UMAP-reduced data.
• This algorithm groups atoms into clusters of arbitrary shapes and densities without requiring a pre-defined number of clusters.
• Output: Atoms are grouped into "ML groups" representing recurring structural motifs (e.g., vacancy cages, dumbbell cores, aggregates).

E. Validation & Quantification

• Cluster Identification (CID): A metric combining the net defect count (interstitials = +1, vacancies = -1) and cluster size is used to physically interpret the ML groups.
• Comparison: The ML results are statistically compared against six conventional methods (CS, CA, DXA, PTM, VOR, Lattice Site Analysis) using Precision, Recall, and F1-scores.

3. Key Results

A. Defect Detection Performance

• The pipeline successfully identified outliers comprising <0.13% of the total atoms in the simulation boxes, consistent with expected residual damage at 80 keV PKA energy.
• >99.7% of identified outliers were assigned to compact, well-defined HDBSCAN groups, demonstrating that almost every anomalous atom belongs to a recurring structural motif.

B. Unsupervised Classification of Defect Types

The ML workflow automatically separated defects into physically meaningful categories without human labeling:

• Vacancy-dominated groups: Identified as cages of first-neighbour atoms surrounding empty lattice sites (monovacancies) and larger vacancy aggregates.
• Interstitial-dominated groups: Identified as dumbbell cores, "halo" atoms surrounding dumbbells, and large interstitial complexes.
• Complex Structures: In Ni, a specific group (Label 3) was identified as surrounding icosahedral (ICO) cores (A15-type structures), capturing the elastically strained shell around the icosahedral kernel which conventional template matching (PTM) missed.

C. Quantitative Correlations

• Size Scaling: A strong correlation ($R^2 > 0.89$) was found between the number of ML-flagged atoms in a cluster and the net defect count ($n_{Def}$). This allows the conversion of atom counts to defect inventories via material-specific calibration curves.
• Material Differences:
  • FeNiCr: Produced larger defect aggregates than pure Ni.
  • Zr (hcp): Showed fewer overall defects but distinct separation between interstitial and vacancy clusters, with interstitials perturbing nearly as many atoms as vacancies (unlike fcc where vacancies are more disruptive).

D. Comparison with Conventional Methods

• High Overlap: The ML approach showed strong overlap with DXA and Centrosymmetry (CS), recovering 70–90% of their outliers with similar precision.
• Complementarity:
  • PTM: The ML workflow captured the "distorted shells" around icosahedral cores that PTM (which relies on rigid templates) missed.
  • CA/VOR: While CA and Voronoi analysis flagged many more atoms (often over-counting), the ML approach provided a tighter, more specific selection of actual structural anomalies.
• No Tuning: The ML workflow achieved this balance of sensitivity and specificity without manual threshold tuning for specific defect types.

4. Significance and Contributions

1. Template-Agnostic Discovery: The method does not require prior knowledge of defect geometries, making it capable of detecting novel or complex defect morphologies (like A15 phases) that traditional template-based methods miss.
2. Unified Framework: It provides a single, consistent pipeline applicable across different crystal structures (fcc and hcp) and chemical compositions (pure metals and alloys).
3. Quantitative Mapping: It transforms qualitative defect detection into a quantitative tool, providing calibration curves to estimate defect counts from atom counts.
4. Scalability: The workflow is designed to handle large-scale MD datasets efficiently, offering a path to analyze massive cascade databases and long-term damage accumulation.
5. Hybrid Potential: The study demonstrates that combining data-driven ML with conventional metrics (like CID or DXA) creates a powerful hybrid toolkit for comprehensive materials characterization.

Conclusion

The paper establishes a robust, unsupervised SOAP $\rightarrow$ Autoencoder $\rightarrow$ UMAP $\rightarrow$ HDBSCAN pipeline for mapping radiation damage. It successfully identifies, classifies, and quantifies defect structures in Ni, FeNiCr, and Zr, outperforming or complementing conventional methods by capturing complex, distorted defect environments that are invisible to standard topological analysis. This framework offers a scalable solution for understanding the atomistic origins of material degradation under irradiation.