Ab Initio Modeling Data Based Autoencoder to Interpret ARPES Data and Assist Inverse Design of Semiconductor Heterostructures

Semiconductor heterostructures instigated tremendous changes in our everyday lives in the form of double-heterostructure lasers for telecommunication systems, heterostructure light-emitting diodes, or high-electron-mobility transistors used in high-frequency devices. Recent advances of nanofabrication techniques have achieved great control over the growth of the heterostructures. Nevertheless, fabrication is strongly affected by the process and the structural properties show high variability. It is essential to acquire a comprehensive understanding of the relationship between growth dependent structural parameters and electronic properties to design semiconductor heterostructures design with desired performance. Ab initio approaches enable prediction of materials properties with minimal experimental input, however, often come with large computational costs. It remains a challenge to model electronic properties of technologically relevant heterostructures incorporating full structural complexity, representing the vast fabrication dependent parameter space. We demonstrated a forward machine learning (ML) approach that can extract the structure-electronic property relationships from the data generated with density functional theory (DFT) calculations and make the first-principles techniques applicable for technologically relevant heterostructures. The ML model can formulate transferable structure-property relationships from DFT data of small 16 atom models and predicts electronic transport coefficients of fabricated semiconductor heterostructures of thickness upto ~12 nm. The key physical understanding revealed by our study is that the relationships between local atomic configurations, and their contributions to global energy bands remain preserved when the local configurations are part of a heterostructures with different composition and/or dimensions. Our study highlights the importance of accurate identification of local structural features on the performance of the ML models. We use neural network (NN) and random forests (RF) algorithms to build the models and compare their performances. The NN performs consistently better than RF. Extending this approach, we exploit the ML models to predict atomically resolved contributions to the effective band structures or the spectral functions of heterostructures. We demonstrate that our ML model can predict such contributions in complex silicon (Si)-germanium (Ge) heterostructures, such as Ge3Si3Ge6Si9Ge15Si24, matching DFT predicted data. Here the numbers refer to the monolayers in the heterostructure. We use this forward learning model to develop an inverse model by implementing a convolutional neural network (CNN) model in the ML workflow. This model receives the effective band structure as input and predict the structural features of the corresponding heterostructure. CNN is a well-established technique for feature extraction in digital images and can assemble more complex patterns from smaller training dataset data samples. As a demonstration, we train the CNN model only with DFT data from heterostructures and task it predicts structural properties of bulk Si. We provide the ARPES image of bulk Si and δ-doped Si as input. The trained CNN model successfully converts the data into a set of structural and atomic features to describe the bulk systems.<br/>Reference: A. Pimachev & S. Neogi, "First-Principles Prediction of Electronic Transport in Fabricated Semiconductor Heterostructures via Physics-Aware Machine Learning" npj Comput Mater 7, 93 (2021)