BOTAN: BOnd TArgeting Network for prediction of slow glassy dynamics by machine learning relative motion

Hayato Shiba, M. Hanai, Toyotaro Suzumura, Takashi Shimokawabe

TL;DR

BOTAN (BOnd TArgeting Network) is a novel Graph Neural Network (GNN) that predicts the slow, complex relaxation of glassy materials by learning both individual particle displacements and the relative motion between neighboring pairs. By "targeting bonds," the model successfully filters out vibrational noise that plagues previous methods, achieving state-of-the-art accuracy in predicting structural rearrangements from a single static snapshot.

What Problem Does This Paper Address?

Predicting the dynamics of a glass—an amorphous solid that looks like a liquid but acts like a solid—is one of the "holy grails" of condensed matter physics. Unlike crystals, glasses have no obvious structural order, yet some regions are much more prone to moving (rearranging) than others.

Recent GNNs have been successful at predicting "particle propensity" (how much a particle is likely to move) based on its local environment. However, these models typically focus only on the motion of individual particles (nodes). This creates a major hurdle: at short time scales, particles vibrate elastically (phononic motion). This vibrational "noise" obscures the actual structural rearrangements the researchers want to predict. Previous models struggled to distinguish between a particle moving because it’s part of a rearrangement and a particle moving because it’s being pushed by a neighbor's vibration.

Methods

The authors introduce BOTAN, an encoder-process-decoder GNN architecture. While it builds on the "Node-Targeting GNN" (NT-GNN) framework, it introduces a critical second target: relative motion.

  • Dual-Target Learning: BOTAN uses two decoders. One predicts the displacement of individual particles (nodes), while the other predicts the change in distance between neighboring pairs (edges/bonds).
  • Message Passing: The model uses 7 iterations of message passing, allowing structural information from distant neighbor shells to propagate to each node and edge.
  • Pretraining Strategy: To handle the vast and complex "loss landscape" of glassy dynamics, the authors use a pretraining phase. They first train the model to predict short-time relative motion (which has a simpler correlation with structure) before moving to the more difficult task of predicting long-term relaxation.
  • Loss Function: A weighted sum of node-based and edge-based $L_2$ losses, controlled by a hyperparameter $p$.

Datasets

The study uses the 3D Kob–Andersen binary Lennard-Jones mixture, a standard "workhorse" model for studying glassy systems.

  • Scale: 4,096 particles per simulation.
  • Ensemble: They utilized the "isoconfigurational ensemble," running 32 separate simulations from the same starting configuration with different initial velocities to calculate the "ground truth" propensity.
  • Volume: 500 initial configurations (400 for training, 100 for testing), resulting in a total of 16,000 simulation runs.

Key Insights and Ideas

  • Edges are the Signal, Nodes are the Noise: Absolute particle displacement is heavily influenced by non-local elastic strains (vibrations). By contrast, the change in distance between neighbors directly characterizes structural rearrangements. By forcing the GNN to learn these "bond" changes, the model learns to "see through" the vibrational noise.
  • The "Metabasin" Shortcut: The authors found that at very short times, there is a simple anti-correlation between initial distance and distance change. Pretraining on this simple relationship helps the model find a "metabasin"—a good starting point in the weight-parameter space—making the training for long-term dynamics much more efficient and accurate.
  • Autonomous Error Correction: BOTAN essentially learns to "interpret" how different dynamical processes (strain vs. rearrangement) affect motion. This allows it to fix prediction errors that node-only models consistently make in regions dominated by collective elastic fluctuations.

Key Results

  • Superior Accuracy: BOTAN consistently outperforms the standard NT-GNN across all temperatures ($T=0.44, 0.50, 0.64$) and all time scales.
  • Short-Time Breakthrough: The most dramatic improvement occurs at short time scales (the $\beta$-relaxation regime). While previous models' accuracy plummeted here due to vibrational noise, BOTAN maintains high Pearson correlation coefficients ($\rho \approx 0.9$ at short times).
  • Spatial Heterogeneity: Qualitative error maps show that BOTAN captures the "hotspots" of movement with much higher spatial precision than previous models, which tended to produce blurred or misaligned predictions in high-strain areas.

Limitations and Future Work

  • Computational Overhead: Calculating and updating features for every pair of neighboring particles (edges) is more computationally expensive than node-only updates, particularly regarding memory and interparticle distance calculations.
  • Static Limitation: Like its predecessors, BOTAN still relies on a single static snapshot. It does not yet incorporate the full history of the system or "reaction coordinates" that might further improve long-term predictions.
  • Architectural Extensions: The authors suggest that future iterations could benefit from more advanced GNN features, such as rotational-equivariant descriptors (to better handle 3D symmetry) or self-attention layers to weight neighbor interactions more intelligently.

Summary updated

Abstract

Recent developments in machine learning have enabled accurate predictions of the dynamics of slow structural relaxation in glass-forming systems. However, existing machine-learning models for these tasks are mostly designed such that they learn a single dynamic quantity and relate it to the structural features of glassy liquids. In this study, we propose a graph neural network model, ``BOnd TArgeting Network (BOTAN)'', that learns relative motion between neighboring pairs of particles, in addition to the self-motion of particles. By relating the structural features to these two different dynamical variables, the model autonomously acquires the ability to discern how different dynamical processes, strain fluctuations and particle rearrangements, affect the self-motion of particles undergoing slow relaxation, and thus can predict with high precision how slow structural relaxation develops in space and time.
BibTeX
@article{shiba2023botan,
  author = {Hayato Shiba and M. Hanai and Toyotaro Suzumura and Takashi Shimokawabe},
  title = {{BOTAN: BOnd TArgeting Network for prediction of slow glassy dynamics by machine learning relative motion}},
  journal = {The Journal of Chemical Physics},
  year = {2023},
  doi = {10.1063/5.0129791},
}
Figure