"Studying Defect Properties in Semiconductors using First Principles Simulations and Machine Learning"

Md Habibur Rahman, MSE PhD Candidate 

Advisor: Professor Arun Kumar Mannodi Kanakkithodi

WebEx Link

ABSTRACT

Every crystal contains imperfections, known as defects, commonly seen as missing, excessive, or misplaced atoms, which play a crucial role in determining equilibrium conductivity that directly affect the optoelectronic properties of semiconductors. These defects could potentially introduce energy states within the bandgap, serving as both traps for charge carriers and potential pathways for photon-emitting transitions. However, pinpointing specific defects through experimental methods can be costly, time-consuming, and uncertain, especially for analyzing a wide range of defects from a large chemical space. As a result, researchers often turn to first principles-based density functional theory (DFT) simulations to estimate the formation energy of a defect as a function of Fermi level (E𝐹) and chemical potential conditions. This, in turn, aids in identifying whether a defect acts as a donor/acceptor, as well as its deep/shallow-level nature, and corresponding equilibrium conductivity of the semiconductor. However, DFT simulations can incur significant computational costs, especially when dealing with complex systems and extensive datasets. These costs result from various factors, such as the necessity for employing large supercells to accurately model the system, the incorporation of charge states to simulate charged defects, and the utilization of advanced levels of theory to get reliable estimation. In contrast, machine learning (ML) models can learn patterns and correlations from large datasets of defects, energies, and properties of interest to discover any potential low energy defects much faster than traditional quantum mechanical simulations. This report demonstrates the potential of advanced ML techniques, like crystal graph-based neural networks, to achieve more precise predictions of defect formation energy compared to traditional ML algorithms.