What do deep neural networks find in disordered structures of glasses?

Abstract

Glass transitions are widely observed in various types of soft matter systems. However, the physical mechanism of these transitions remains elusive despite years of ambitious research. In particular, an important unanswered question is whether the glass transition is accompanied by a divergence of the correlation lengths of the characteristic static structures. In this study, we develop a deep-neural-network-based method that is used to extract the characteristic local meso-structures solely from instantaneous particle configurations without any information about the dynamics. We first train a neural network to classify configurations of liquids and glasses correctly. Then, we obtain the characteristic structures by quantifying the grounds for the decisions made by the network using Gradient-weighted Class Activation Mapping (Grad-CAM). We consider two qualitatively different glass-forming binary systems, and through comparisons with several established structural indicators, we demonstrate that our system can be used to identify characteristic structures that depend on the details of the systems. Moreover, the extracted structures are remarkably correlated with the non-equilibrium aging dynamics in thermal fluctuations.

Dead or alive: Distinguishing active from passive particles using supervised learning (a)

“…structural indicators that predict dynamical properties in densely disordered (near-)equilbrium systems [6][7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25]. These findings firmly establish a correlation between local structure and the propensity of passive particles to move in a crowded environment.…”

Rotation-equivariant graph neural networks for learning glassy liquids representations

“…• Use coarse-grained mobility measures as target labels. Indeed, they have been shown to be more structure-dependent [50] and thus also display larger correlation coefficients with simple structural descriptors [67]. Eliminating noise in the labels could help achieve better precision, possibly reaching the maximum.…”

Machine Learning in Soft Matter: From Simulations to Experiments

“…Nonetheless, when the number of features is high and the underlying relationship is complex, these models tend to lose its white-box characteristics, primarily due to the challenging nature of comprehending its decision-making process. To understood more complex models, post hoc interpretation methods, such as SHAP, [343] LIME, [344] and Grad-CAM, [345,346] are employed to offer insights into the learned knowledge of well-trained blackbox models without altering these models. [347] These methods not only provide explanations for individual predictions made by models, such as what descriptors and/or structures led to the particular prediction, [249] but also reveal the global relationships that the model has learned, such as what descriptors are associated with prediction by using importance scores of descriptors.…”