Grainger College of Engineering, All Events

Entropy is a unifying concept that lies at the heart of both statistical mechanics and information theory. In this talk, I will illustrate how these two notions of entropy—physical and informational—can play central and often underappreciated roles in modern computational materials science.

I will begin with a recent study of dislocation nucleation kinetics at free surfaces in metals, where molecular dynamics simulations revealed strikingly anomalous behavior. By combining an analysis of vibrational entropy along the nucleation pathway with a variational rate theory, we show that both the unusual kinetic prefactors and their strongly non-Arrhenius temperature dependence emerge naturally. This work highlights how standard physical intuitions can fail dramatically when entropy is not treated as a first-class contributor to kinetics.

In a second vignette, I will turn to information theory and its application to machine-learning models of materials. I will show how entropy-based principles provide a systematic and physically grounded framework for dataset construction—a process that has traditionally relied heavily on heuristic choices. By maximizing information entropy over an ML feature space, we obtain exceptionally robust models with minimal human intervention, enabling the development of truly “universal” machine-learning potentials that remain accurate across unprecedented physical and chemical spaces.