University of Colorado – Boulder, USA
From Atoms to Devices: Bridging the Gap with Physics-Informed Machine Learning Models
Abstract: First-principles techniques for calculating phonon and electron transport properties of materials have progressed rapidly in recent years. These techniques enable prediction of materials properties with minimal experimental input, however, often come with large computational costs. For example, the calculations of electronic transport coefficients (such as, thermopower or conductivity) require large number of individual energy calculations and computational costs can accrue quickly. On the other hand, broad spectra of phonon modes and diverse scattering mechanisms, makes prediction of phonon transport properties, especially of complex nanostructures, via first-principles models particularly challenging. As a result, the applicability of these techniques remains limited to model electronic or thermal properties of technology enabling heterostructures incorporating full structural complexity, representing the vast fabrication dependent parameter space. Alternatively, the rich electron or phonon dynamics in these spaces makes the problem ideal for the application of data-driven techniques. Machine-learning-based-materials-informatics (MI) approaches are increasingly used to accelerate design and discovery of new materials with targeted properties, however, few studies exploited MI to learn the atomic scale dynamics or physics of carrier transport in complex systems. We develop physics-informed machine learning (ML) approaches to predict electron and phonon transport properties of nanostructures. The physics input allows us to limit the size of training data and develop ML model descriptors that are strongly correlated with the transport properties and thus expected to have a strong role on predictability. Our electronic-transport-informatics (ETI) framework is built on the hypothesis that functional relationships between local atomic configurations, and their contributions to global energy states, are preserved when the local configurations are part of a larger system. The framework is trained on the electronic structure properties of small systems and predicts transport coefficients, namely the thermopowers of experimental semiconductor heterostructures. Our ETI framework thus demonstrates that MI can be exploited to address the gap between ideal first-principles accessible models and systems realized with nanofabrication techniques.
Sanghamitra Neogi is an Assistant Professor in the Smead Aerospace Engineering Sciences Department at the University of Colorado Boulder since Fall 2015. Prior to joining UC-Boulder, she received her Ph.D. in theoretical condensed matter physics from the Pennsylvania State University and was a postdoctoral research associate at the Max Planck Institute for Polymer Research, Mainz, Germany. Her research focuses on characterizing structure-processing-property relationships in nanostructured materials, phonon engineered nanomaterials, thermoelectric energy conversion, application of machine learning methods to predict heat and charge transport in nanostructures, harnessing stochastic phonon processes for novel computing applications, and probing spin-phonon interaction in solid state quantum systems.