What Do Deep Neural Networks Find in Disordered Structures of Glasses?
Speaker
Takeshi Kawasaki
The University of Osaka
Abstract

Glass transitions are widely observed across diverse soft matter systems. Yet despite decades of extensive research, the physical mechanisms underlying these transitions remain elusive. A central open question is whether the glass transition is associated with a diverging correlation length of characteristic static structures [1,2].

In this study, we develop a deep learning-based method to extract characteristic local structures directly from instantaneous particle configurations, without recourse to any dynamical information. We train a convolutional neural network (CNN) to classify configurations as either liquid or glass, and then apply gradient-weighted class activation mapping (Grad-CAM) to identify the structural features driving this classification by highlighting the regions most influential in the network's decision-making. Applying this framework to two model glass-forming liquids and comparing the results against several established structural indicators [2], we confirm that our method successfully captures system-specific structural features. We further find that the extracted structures exhibit strong correlations with long-time particle displacements [3,4].

From a methodological standpoint, earlier CNN-based approaches [3,4] and conventional structural indicators [2] alike required manual coarse-graining of the structural indices obtained. Our updated method addresses this limitation by employing a graph neural network (GNN), which improves computational efficiency and simultaneously enables automatic coarse-graining of structural information [5].  Our results suggest that structural features—hidden within disordered configurations yet revealed through these approaches—play a significant role in the glass transition.

References:

1. T. K., T. Araki, and H. Tanaka, Phys. Rev. Lett. 99, 215701 (2007); H. Tanaka, T. K., H. Shintani, and K. Watanabe, Nat. Mater. 9, 324 (2010).

2. H. Tong and H. Tanaka, Phys. Rev. X 8, 011041 (2018).

3. N. Oyama, S. Koyama, and T. K., Frontiers in Physics 10, 1007861 (2023).

4. M. Liu, N. Oyama, T. K., and H. Mizuno, J. Appl. Phys. 136, 144702 (2024).

5. M. Liu, D. Nishigaito, Y. Hara, N. Oyama, and T. K., to be submitted.
--
FROM 202.120.11.*