Machine-learning integrated glassy defect from an intricate configurational-thermodynamic-dynamic space

Abstract

Optimizing materials' properties and functions by controlling defects in the crystalline phase has been a cornerstone of materials science and condensed matter physics. However, this paradigm has yet to be established in the broadly defined amorphous materials, which implies the identification of very subtle structural features in an otherwise uniformly disordered medium. Here we propose and define a new integrated glassy defect (IGD), based on machine learning strategy informed by atomistic physics, and also by an extremely wide configurational, thermodynamic, and dynamic variables space of the disordered state. The IGD simultaneously includes positional topology and vibrational features, as well as the local morphology of the potential energy landscape. This unprecedented combination gives rise to a much more comprehensive and more effective definition of the ``glassy defect,'' much beyond the conventional, purely structural input. IGD can be used not only as an efficient predictor of athermal plasticity but is also transferable to detect both short-time vibrational anomalies (the boson peak), and long-time relaxation and diffusion dynamics in glasses. The integrated strategy is instrumental to build the long-sought structure-property relationship in complex media.