Data driven insights into the characteristics of wide bandgap semiconductors in 2D materials

[Display omitted] The ever-increasing complexity in the working conditions of power electronics has demanded a search for the semiconductor materials that operate at high temperatures and high voltage conditions. Coupled with the need for smaller devices, this has prompted the study for semiconductors of lower dimensionality. Two-dimensional wide band gap semiconductors fit these criteria well and are potentially ideal candidates. Typically, the search of these materials using conventional domain knowledge is highly time consuming and computationally expensive because of the need to compute the material properties using high-fidelity methods. Recent advances in machine learning methods have allowed for data driven decisions which reduce the burden of significant domain knowledge from the end users. Data driven analysis combined with machine learning algorithms have found remarkable success in the inverse design and search for materials with desired properties. In this work, we perform a data driven study using the C2DB dataset, to explore the relationships between local structural, electronic, and compositional features of 2D materials and their corresponding bandgap values. The core features responsible for wide bandgaps in semiconductor materials are identified via exploratory data analysis. We build two sets of neural networks – a simple artificial neural network and a graph neural network which are then trained to predict the bandgap of 2D materials. The graph neural network is subsequently trained to predict formation energy of the 2D materials.