"Simulating and Understanding Defects in Semiconductors using Density Functional Theory and Machine Learning" 

Habibur Rahman, MSE PhD Candidate 

Advisor: Professor Arun Mannodi Kanakkithodi

ABSTRACT

The theoretical maximum efficiency of CdTe is around 32%, but real devices remain stuck near 22%—a persistent gap driven by point defects in the absorber that act as non-radiative recombination centers. Similar defect-limited gaps plague every thin-film photovoltaic (PV) technology, yet predicting defect formation energies, charge transition levels, and equilibrium concentrations remains computationally prohibitive at the accuracy required (hybrid functionals with spin–orbit coupling). Traditional density functional theory (DFT) workflows also suffer from metastable local minima, limited transferability across chemistries, and poor scaling. This dissertation addresses six barriers to high-throughput defect engineering in chalcogenide semiconductors. (1) Standard DFT relaxations had long trapped defects in metastable minima, causing transition levels to disagree with experiment by tenths of an eV; we applied systematic symmetry-breaking across our chalcogenide library to recover ground-state geometries and resolve these discrepancies. (2) The role of defect complexes and extrinsic dopants was fragmented in the literature; we now map how native defects, substitutional dopants and impurities (As, Cl, Cu, O, other group-V), and their complexes across II–VI chalcogenides reshape formation energies, equilibrium Fermi levels, and carrier concentrations, explaining why certain compositions outperform others. (3) Existing machine learning models could not handle charged defects, multiple levels of theory, or prediction uncertainty in a unified way, and crucially lacked any way to flag when their predictions were trustworthy; we built a graph neural network force field that takes defect structure, charge state, and level of theory as inputs and predicts energies, forces, and transition levels together with a calibrated uncertainty estimate, allowing users to identify regions of chemical space where the model can be trusted and where additional DFT data are needed. (4) Predicted defect structures were rarely validated against experiment; our simulated geometries now fit measured X-ray absorption near-edge structure (XANES) spectra exceptionally well, confirming predictive accuracy. (5) Computational screening efforts rarely considered defect tolerance and dopability; our machine-learning-accelerated screening of half a million chalcogenide compositions identifies novel defect-tolerant candidates. (6) Defect workflows required high-performance computing (HPC) expertise; we released open-source tools on the nanoHUB platform giving interactive access to our databases and automated workflows. Together, these contributions deliver an end-to-end pipeline from first-principles defect physics to community-accessible design tools for next-generation photovoltaic, electronics, and quantum devices.