Tailoring Semiconductor Defect Properties Using Graph Neural Networks and Active Learning

Defects and impurities in semiconductors significantly impact their performance in solar cells, photocatalysis, electronic devices, and related applications [1,2]. To address the challenge of quick defect property prediction in semiconductors and subsequent design of novel materials with tailored defect behavior, we developed a comprehensive workflow that integrates active learning (AL) and crystal graph-based neural network (GNNs) models trained on high-throughput density functional theory (DFT) data [2,3]. We generated a large DFT dataset of the crystal structures of technologically-important semiconductors, including well-known binary compounds such as CdTe, ZnS, and SiC, a variety of anion-site or cation-site alloys, and many types of vacancies, self-interstitials, anti-site substitutions, extrinsic interstitial and substitutional defects, and defect complexes simulated in them. Using an innovative approach of sampling optimized, partially optimized, and unoptimized configurations from DFT calculations, we rigorously trained GNN models on &gt; 20,000 crystal structures to yield accurate prediction of the crystal formation energy (CFE) corresponding to any bulk or defective structure in multiple charge states [3].<br/><br/>Using the Atomistic Line Graph Neural Network (ALIGNN) framework [4], we obtain test root mean square errors (RMSE) of &lt; 5 meV/atom for the CFE, which significantly outperforms the current state-of-the-art and is a remarkable accuracy considering its generalizability to supercell sizes, type of alloying, and missing or additional atoms. We also combine ALIGNN predictions with a systematic distortion-based geometry optimization approach which provides an effective surrogate for DFT computations to obtain low energy defective crystal structures for any semiconductor-defect combination [3]. Using AL, model predictions are iteratively improved by launching new DFT calculations based on prediction uncertainties and retraining the models until errors converge. The CFE predictions are extended to obtain defect formation energy plots as a function of Fermi level, chemical potential conditions, and defect charge, and used to screen through thousands of possible defects and dopants (e.g., group V dopants, Cu, and unintentional impurities like Cl, O, and associated complexes) and generate libraries of low energy defect configurations in the (Cd,Zn)-(Te,Se,S) compositional space, relevant for CdTe-related solar cells [5,6]. These defect configurations are utilized to interpret experimental spectroscopy measurements [7,8] and devise ways to improve optoelectronic performance by tuning semiconductor composition, eliminating certain impurities, and tailoring chemical growth conditions.<br/><br/><b>REFERENCES</b><br/>[1] A. Mannodi-Kanakkithodi, “A guide to discovering next-generation semiconductor materials using atomistic simulations and machine learning,” Comput. Mater. Sci. 243, 113108–113108 (2024).<br/>[2] A. Mannodi-Kanakkithodi et al., “Universal machine learning framework for defect predictions in zinc blende semiconductors”, Patterns. 3, 100450 (2022).<br/>[3] M. H. Rahman <i>et al.</i>, “Accelerating defect predictions in semiconductors using graph neural networks,” <i>APL Machine Learning</i>, 2, 0166122 (2024).<br/>[4] Choudhary, K. et al. Atomistic Line Graph Neural Network for improved materials property predictions. NPJ Comput. Mater. 7, 185 (2021).<br/>[5] M. H. Rahman et al., “First Principles Investigation of Dopants and Defect Complexes in CdSe_xTe_1-x”, in preparation.<br/>[6] M. H. Rahman et al., “Discovering Low Energy Defect Configurations in Cd/Zn-Te/Se/S Compounds using Graph Neural Networks and Active Learning”, in preparation.<br/>[7] S. Rojsatien et al., “Quantitative analysis of Cu XANES spectra using linear combination fitting of binary mixtures simulated by FEFF9”, Rad. Phys. Chem. 202, 110548 (2022).<br/>[8] S. Rojsatien et al., “Distribution of Copper states, phases and defects across the depth of a Cu-doped CdTe solar cell”, Chem. Mater. 35, 23, 9935-9944 (2023).