Inferring the dynamics of glass-forming liquids from static structure across thermal states

Abstract

In this study, we demonstrate the generalizability of graph neural networks in predicting the dynamic heterogeneity of model glass-forming liquids across different temperatures. While previous approaches have often been limited to making predictions at the specific temperatures used during training, we find that our proposed framework - T-BOTAN - enables interpolation to temperatures not included in the training set. We show that the dynamical behavior, the associated four-point correlations, and even the macroscopic temperature can be estimated with sufficient accuracy solely from static particle configurations at untrained temperatures. These results suggest that static configurations encode not only local structural features driving dynamic heterogeneity but also fundamental thermodynamic information.

Figures (15)

• Figure 1: Schematic representation of the T-BOTAN architecture. Here, "EN", "DE", and "MLP" denote the encoder, decoder, and multilayer perceptron, respectively. Node features encode particle species, while edge features consist of the relative particle coordinates in the inherent structure and the interparticle distances in the thermal structure. Message passing iteratively updates node, edge, and global features, followed by node-, edge-, and global-level output heads that predict bond-breaking correlations, interparticle distance changes in the thermal structure, and temperature.
• Figure 2: Comparison of MD simulation data ((a)-(d)) and T-BOTAN predictions ((e)-(h)) for the spatial distribution of $C_{\mathrm{B}}(t=\tau_{\alpha})$ at $T=0.44, 0.47, 0.50$, and $0.56$. All data are plotted for the same three-dimensional particle configuration, in the cross section $11.1<z<11.9$.
• Figure 3: Comparison of $S_{\mathrm{B}}(q,t=\tau_{\alpha})$ from MD simulation data with the prediction of T-BOTAN for various $T$. (a) $T=0.44$ (b) $T=0.47$ (c) $T=0.50$ (d) $T=0.56$
• Figure 4: Comparison of $T_{\mathrm{pred}}$ with $T_{\mathrm{MD}}$. Black symbols denote temperatures used as training data, and red symbols denote temperatures not used for training. Error bars represent the standard deviation over different samples, but are smaller than the symbol size and thus not visible. The dashed line indicates $T_{\mathrm{pred}}=T_{\mathrm{MD}}$.
• Figure 5: Pairwise distances $D(T_i,T_j)$ between mean global latent representations learned by T-BOTAN at different temperatures. For each temperature, global embeddings obtained at the final message-passing layer (before decoding) are averaged over configurations and random seeds. Small distances indicate that the network encodes configurations at the two temperatures in a similar manner, revealing that temperatures form discrete clusters in the learned feature space. The diagonal components are by definition zero.