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.
What problem does this paper attempt to address?
### Problems the paper attempts to solve
This paper aims to solve the problem of how structural factors affect kinetic behavior in amorphous solids. Specifically, the research focuses on improving the interpretability of models in glassy state research by introducing a new machine - learning method - Structural Constraint Neural Network (SCNN). Traditional machine - learning models can usually only predict kinetic properties from structural inputs, while SCNN can not only predict kinetic properties but also learn structural prototypes under different kinetic patterns during the training process. These learned structural constraints can be used as a key breakthrough point to explain how structural differences affect kinetics.
### Background and motivation
Amorphous alloys (metallic glasses) are a class of disordered materials with unique local atomic arrangements. Unlike crystalline materials, metallic glasses lack translational and rotational periodicity, which makes defect detection more complex. Although machine - learning techniques have made progress in recent years in identifying structures that affect kinetic properties, many methods still lack transparency, affecting the interpretability of the models. Therefore, finding a balance between performance and interpretability remains an unsolved problem in glassy - state physics.
### Research methods
1. **Dataset description**:
- Use the activation energy data of Cu - Zr metallic glass samples for testing.
- The dataset contains six Cu - Zr metallic glass models with different cooling rates and compositions.
- Obtain single - atom activation energy through molecular dynamics simulation (MD) and activation - relaxation technique (ARTn).
2. **Structural descriptors**:
- The structural parameters input into the model include radial density and orientational order parameter.
- The radial density is calculated by the Bessel basis function, and the orientational order parameter is calculated by the spherical harmonics.
3. **Structural Constraint Neural Network (SCNN)**:
- SCNN consists of a feature embedding layer, a structure extraction layer, and a classification layer.
- The feature embedding layer uses a multi - layer perceptron (MLP) to reduce the 264 - dimensional structural parameters to 32 - dimensional.
- The structure extraction layer learns structural constraints and calculates the similarity between each particle and the structural constraints.
- The classification layer classifies according to the similarity score.
4. **Training process**:
- Randomly initialize the structural constraints in the initial stage.
- Train the model by optimizing the loss function, which includes cross - entropy loss, clustering loss, separation loss, orthogonal loss, and structural separation loss.
- Update the structural constraints through a replacement strategy to align them with specific particles, enhancing transparency and traceability.
- Finally, optimize the weight matrix of the classification layer.
### Experimental results
1. **Predictive ability**:
- The AUC - ROC value of the SCNN model on the test dataset is 0.869. Compared with the traditional multi - layer perceptron (MLP) model (AUC - ROC value of 0.882), there is no significant performance degradation.
- The calibration curve shows that the predicted probability of the SCNN model is highly correlated with the proportion of actual high - activation - energy atoms.
2. **Structural constraint extraction**:
- By reducing the particle feature embedding to a two - dimensional space through principal component analysis (PCA), it is found that high - activation - energy and low - activation - energy atoms are well - clustered in the feature space.
- The structural constraints are located at the centers of their respective categories, and there is a high similarity between the constraints within the category, and there are significant differences between the constraints between the categories.
3. **Relationship between structural constraints and activation energy**:
- The structural characteristics of low - activation - energy atoms are similar to the first two structural constraints, while the structural characteristics of high - activation - energy atoms are similar to the last two structural constraints.
- The effectiveness of the structural constraints is verified by clustering all atoms, and atoms in different clusters show different dynamic characteristics.
4. **Analysis of structural constraints**:
- The local atomic structures of representative particles are visualized through the Largest Standard Cluster Analysis (LaSCA), and it is found that there are significant structural differences in the atomic environments corresponding to different structural constraints.
- These structural differences reveal the main differences between solid - state and liquid - state structures, providing a new perspective for understanding the structure - kinetics relationship.
### Summary
This paper, by introducing the Structural Constraint Neural Network...