Comparing machine learning techniques for predicting glassy dynamics

Abstract

In the quest to understand how structure and dynamics are connected in glasses, a number of machine learning based methods have been developed that predict dynamics in supercooled liquids. These methods include both increasingly complex machine learning techniques, and increasingly sophisticated descriptors used to describe the environment around particles. In many cases, both the chosen machine learning technique and choice of structural descriptors are varied simultaneously, making it hard to quantitatively compare the performance of different machine learning approaches. Here, we use three different machine learning algorithms -- linear regression, neural networks, and GNNs -- to predict the dynamic propensity of a glassy binary hard-sphere mixture using as structural input a recursive set of order parameters recently introduced by Boattini et al. [Phys. Rev. Lett. 127, 088007 (2021)]. As we show, when these advanced descriptors are used, all three methods predict the dynamics with nearly equal accuracy. However, the linear regression is orders of magnitude faster to train making it by far the method of choice.

Neural Networks Reveal the Impact of the Vibrational Dynamics in the Prediction of the Long-Time Mobility of Molecular Glassformers

“…First, let us focus on the NN B , i.e., the one considering only the structural descriptors discussed in Section 2.2.1 and not resorting to information about the local DW (right panel). The correlation agrees with the reported literature and exhibits a well-defined peak at [ 13 , 14 , 22 ]. In particular, at very short times ( ) the correlation is rather small and even slightly less than the one in hard-sphere systems [ 13 ].…”

“…The correlation agrees with the reported literature and exhibits a well-defined peak at [ 13 , 14 , 22 ]. In particular, at very short times ( ) the correlation is rather small and even slightly less than the one in hard-sphere systems [ 13 ]. At intermediate times, at the onset of the cage regime ( ), corresponding to the start of the quasi-plateau region observed in the left panel of Figure 2 , we observe a minimum of the Pearson correlation.…”

“…This means that the prediction of the propensity at all the investigated times happens simultaneously. This approach is different from the one adopted in previous works where the prediction is made one time interval at a time [ 13 , 14 , 22 ]. Such a choice has been made to force the network to produce a model that works at all the time-scales.…”

“…ML strategies in the field of the glass-forming liquid focus on predicting the dynamical behavior based on local structural features. Starting from the pioneering work by Cubuk et al [ 12 ] relying on support vector machines (SVMs) to predict rearrangement probability in glassy mixtures, more and more cutting edge techniques have been employed to accomplish similar tasks [ 13 , 14 , 15 , 16 , 17 , 18 ].…”

“…However, the quality of the predictions achieved by the GNNs developed by both by Bapst et al and Boattini et al, even if remarkable, is not perfect and motivates further developments. Most important, the above mentioned studies investigated only atomic liquids (single component or mixtures) [ 13 , 14 , 22 ]. Thus, the performance of the GNNs when tackling connected systems such as molecular liquids is unclear.…”