Date and Time: 22nd July 2022 (Friday) Time: 2:15 PM Tea/Coffee: 02:00 PM
Venue: Kanada Auditorium, JNCASR
Title: Learning the Dynamics of Disordered Systems with Graph Neural Networks
Abstract:
Understanding the structure–dynamics correlation in disordered systems such as liquids and glasses has been an open problem of both fundamental importance and practical relevance. For instance, although glasses have been used commercially for more than 2 millennia, the phenomenon of glass transition still remains elusive to scientists. Recent advances in data-driven approaches such as machine learning present a promising method to model and learn these systems. Especially, graph neural networks (GNNs), which infuses an additional inductive bias in the form of the topology of the structure, presents as an ideal approach to learn the structure–dynamics correlation. In this talk, we will discuss two approaches employing GNNs which can be used for deciphering disordered systems. The first focusses on employing unsupervised GNNs to predict the local structural motifs in disordered systems. We show that the local motifs predicted by the GNNs correspond to low and high energy clusters. Further, we demonstrate the dynamics of these clusters are directly proportional to their energy, thereby providing insights into the structure—dynamics correlation. The second approach focusses on learning the interaction laws in disordered systems directly from their trajectory. To this extent, we propose three different physics-informed GNNs, namely, Lagrangian and Hamiltonian GNNs, and Graph Neural ODEs. By applying these GNNs to different systems, namely, spring system, LJ system, and silica, we show that the interaction laws governing the dynamics of the atoms can be learned directly from the trajectory with no a priori knowledge on the forces or the energy of the system. Finally, we will briefly discuss some future directions on how the performance of these GNNs could be improved and tailored to solve some of the fundamental problems in glass science.
