Deep Learning Model for Inverse Design of Semiconductor Heterostructures with Desired Electronic Band Structures

Abstract

First-principles modeling techniques have shown remarkable success in predicting electronic band structures of materials. However, the computational costs make it challenging to use them for predicting band structures of semiconductor heterostructures, that show high variability of atomic structures. We propose a machine learning-assisted first-principles framework that bypasses expensive computations and predicts band structures from the knowledge of atomic structural features. Additionally, the framework directly connects modeling results and experimental data. For example, it accepts images obtained with angle-resolved photoemission spectroscopy as input and predicts the corresponding atomic structures. The framework leverages the physical relationship between atomic environments and bands. We demonstrate the framework using silicon/germanium based superlattices and heterostructures. Once trained on silicon-based systems, the framework can even predict band structures of gallium arsenide thin films. The physics-informed framework establishes an approach to expedite the design and discovery of complex materials with desired band structures, going beyond combinatorial approaches.