Towards interpreting the thermally activated $β$ dynamics in metallic glass with the structural constraint neural network

Jiang Xiao, Zean Tian, Kenli Li, Wangyu Hu

TL;DR

Researchers have developed the Structural Constraint Neural Network (SCNN), a model that doesn't just predict how atoms move in metallic glasses, but explains why by identifying physical "prototypes." By forcing the network to compare every atom to specific, real-world structural benchmarks, the team successfully linked high-energy atomic stability to 5-fold symmetry (icosahedral packing) and identified the specific structural defects that trigger liquid-like behavior.

What Problem Does This Paper Address?

Metallic glasses (MGs) are a material scientist's riddle. Unlike crystalline metals, which have a neat, repeating lattice where "defects" are easy to spot, MGs are amorphous—their atoms are jammed together in a disordered, glass-like state.

This lack of order makes it incredibly difficult to predict $\beta$ dynamics (the thermally activated local rearrangements of atoms). We know some atoms are "solid-like" (stuck in place) while others are "liquid-like" (prone to moving), but we don't fully understand the structural "defects" that cause this difference. While deep learning models can predict these dynamics with high accuracy, they are usually "black boxes." They tell you what will happen, but not why the structure leads to that outcome. This paper aims to bridge that gap by creating a model that is both high-performing and physically interpretable.

Methods

The authors introduce the Structural Constraint Neural Network (SCNN). Its architecture is inspired by "Prototypical Neural Networks" used in image recognition (e.g., a model that learns what a "typical" wing looks like to identify a bird).

  1. Structural Descriptors: Instead of raw coordinates, the model uses hand-crafted features: Bessel basis functions for radial density (more efficient than standard Gaussian functions) and spherical harmonics for orientation order. They also use shell-averaging to capture how an atom's neighbors' neighbors influence its stability.
  2. The Prototype Layer: The core innovation is the "structure extraction layer." During training, the model learns four "structural constraints" (prototypes)—two representing low-activation energy (liquid-like) environments and two representing high-activation energy (solid-like) environments.
  3. The Replacement Strategy: To ensure these prototypes are physically meaningful, the authors implemented a strategy where the abstract mathematical prototypes are replaced by the embeddings of actual particles from the training set. This forces the model to base its decisions on real, physical atomic configurations.
  4. Similarity-Based Prediction: The final classification is based on how similar an atom is to these four physical prototypes using cosine distance.

Datasets

The study utilized a benchmark dataset of Cu-Zr (Copper-Zirconium) metallic glass models.

  • Source: Molecular Dynamics (MD) simulations.
  • Scale: Six different models with varying cooling rates and compositions, totaling 24,000 particles.
  • Labels: The "ground truth" was determined using the Activation-Relaxation Technique nouveau (ARTn), which calculates the actual energy barrier (activation energy, $E_{act}$) an atom must overcome to move.

Key Insights and Ideas

  • Interpretability doesn't require a performance hit: The SCNN achieved an AUC-ROC of 0.869, which is nearly as high as a standard "black box" Multi-Layer Perceptron (0.882). This proves that adding physical constraints doesn't have to ruin the model's predictive power.
  • 5-Fold Symmetry is the Key to Stability: The model's "solid-like" prototypes ($S_2$ and $S_3$) were found to be rich in 5-fold symmetry (specifically Z12 icosahedral packing). This confirms a long-standing hypothesis in glassy physics: dense, symmetric packing creates high energy barriers that prevent atoms from moving.
  • Identifying "Liquid-Like" Defects: Conversely, the "liquid-like" prototypes ($S_0$ and $S_1$) showed a high prevalence of structural "defects" like S444 and S666 clusters. These are lower-density arrangements that provide the "room" for atoms to rearrange under thermal excitation.
  • Data-Driven Discovery of Physics: By using the "replacement strategy," the researchers moved from abstract machine learning to "data-driven physics," where the model points to specific atoms and says, "This is what a stable environment looks like."

Key Results

  • Predictive Accuracy: The SCNN accurately classified high vs. low activation energy atoms with an AUC of 0.869.
  • Structural Correlation: Using Largest Standard Cluster Analysis (LaSCA), the authors verified that the atoms the model chose as prototypes were indeed topographically distinct. For example, prototype $S_2$ was a perfect icosahedron ([12/S555]), while $S_0$ was a much more disordered cluster.
  • Generalization: The authors showed that even using the raw physical descriptors of these four prototypes (bypassing the neural network's complex embedding) could still classify atoms with an AUC of 0.767, proving the prototypes captured the essential physics of the system.

Limitations and Future Work

  • Feature Engineering: The model still relies on hand-crafted structural descriptors. The authors suggest that moving toward Graph Neural Networks (GNNs) could allow the model to learn its own descriptors from raw atomic positions while maintaining the prototype-based interpretability.
  • Scope: The study focused specifically on Cu-Zr alloys and $\beta$ dynamics. Whether these specific structural prototypes hold true for other types of metallic glasses or different dynamic regimes (like $\alpha$ relaxation) remains to be seen.
  • Medium-Range Order: While the model uses shell-averaging, it primarily looks at local environments. Future work could explore even larger-scale "medium-range order" to see how clusters of clusters influence material failure or aging.

Summary updated

Abstract

Unraveling the structural factors influencing the dynamics of amorphous solids is crucial. While deep learning aids in navigating these complexities, transparency issues persist. Inspired by the successful application of prototype neural networks in the field of image analysis, this study introduces a new machine-learning approach to tackle the interpretability challenges faced in glassy research. Distinguishing from traditional machine learning models that only predict dynamics from the structural input, the adapted neural network additionally tries to learn structural prototypes under various dynamic patterns in the training phase. Such learned structural constraints can serve as a breakthrough in explaining how structural differences impact dynamics. We further use the proposed model to explore the correlation between the local structure and activation energy in the CuZr metallic glass. Building upon this interpretable model, we demonstrated significant structural differences among particles with distinct activation energies. The insights gained from this analysis serve as a data-driven solution for unraveling the origins of the structural heterogeneity in amorphous alloys, offering a valuable contribution to the understanding the amorphous materials.
BibTeX
@article{xiao2024towards,
  author = {Jiang Xiao and Zean Tian and Kenli Li and Wangyu Hu},
  title = {{Towards interpreting the thermally activated $β$ dynamics in metallic glass with the structural constraint neural network}},
  journal = {The Journal of Physical Chemistry Letters},
  year = {2024},
  doi = {10.1021/acs.jpclett.4c00280},
}
Figure