Direct and Reverse Structure-Electronic Property Relationship Prediction with Deep Learning and Bayesian Optimization

Ab-initio computational approaches fail to predict electronic properties of systems with increasing size and complexity. It is highly desirable to formulate methods that can translate the information from ab-initio techniques across length scales, and predict electronic properties of technological applications at relevant length scales. We use a deep learning model to find the complex relationship between the geometrical attributes of heterostructures (HS) and the electronic structure properties predicted by ab-initio calculations. We test the usefulness of the model by predicting electronic transport properties of unknown Si/Ge HS based on the relationship, and comparing with experimental data. We consider the local atomic environment features and the global HS features as the descriptors of the model. The advantage of deep learning models becomes evident since the configuration space is vast due to the variability of HS fabrication. We propose a reverse approach based on Bayesian optimization to predict the structure from measured system’s properties of interest. This method is generically applicable to predict electronic properties of dynamically varying heterostructures and the discovery of new structures.