Unified Graph Neural Network Model to Accelerate Substitution and Defect Engineering in Halide Perovskites

ABX₃ halide perovskites have been the subject of extensive investigation over the last two decades due to their attractiveness for single-junction and tandem solar cells, light emission, sensors, and photocatalysis. The flexible structure of perovskites and their ability to accommodate both inorganic and organic cations provide a broad spectrum of options for tuning their electrical and optical properties [1]. Despite numerous data-driven efforts, the massive chemical space of perovskites when considering cation/anion-site alloying, point defects, and dopants, combined with the expense of running high accuracy density functional theory (DFT) computations, pose a significant challenge in identifying low energy bulk and defect configurations and designing novel atom-composition-structure (ACS) [2] combinations with multiple desired properties.<br/>We present an autonomous crystal structure prediction (CSP) workflow based on crystal graph-based neural networks (GNNs), where starting from any ABX₃ composition, low energy configurations are devised by identifying suitably alloying elements and likely point defects or complexes. Predictive GNN models are trained on a dataset of &gt; 20,000 perovskite structures generated by our group over the years [3,4], spanning a variety of A/B/X species, different phases and supercell sizes, substitutional alloys, lattice strains and octahedral distortions, native point defects, and dopants from across the periodic table. Applying established methods such as CGCNN [5], M3GNET [6], and ALIGNN [7], we obtain crystal formation energy prediction errors &lt; 13 meV/atom, which is competitive with the state-of-the-art. These predictions are combined with systematic distortion-based geometry optimization [8] (e.g., using the MatGL approach [6]) to design new stable perovskite alloy structures and obtain a complete picture of charge-dependent defect formation energies in them, which further informs their defect tolerance, n-type or p-type nature, and likelihood to be doped a certain way. Furthermore, we incorporate the PU learning approach [9] within our GNN models to predict the likely synthesizability of new compounds, utilizing the experimental data we collected from the literature over the years [4]. Combined with separate predictions of electronic band gap and photovoltaic efficiency [4] as well as systematic hybrid functional calculations, we ultimately discover novel perovskite ACS combinations with desired bulk stability and likelihood of synthesis, defect tolerance and dopability, and suitable optoelectronic properties. This framework will continue to grow with new data and improved models and promises to lead perovskite discovery for the next few years.<br/><b>References:</b><br/>[1] J. Yang et al., <i>MRS Bulletin.</i> 47, 940–948 (2022).<br/>[2] A. Zunger, <i>Nat Rev Chem.</i> 2, 0121 (2018).<br/>[3] J. Yang et al., <i>Digital Discovery.</i> 2, 856-870 (2023).<br/>[4] J. Yang et al., <i>J. Chem. Phys. </i>160, 064114 (2024).<br/>[5] T. Xian et al., <i>Phys. </i><i>Rev. Lett. </i>120, 145301 (2018).<br/>[6] C. Chen et al., <i>Nat. Comput. Scie. </i>2, 718 (2023).<br/>[7] K. Choudhary et al., <i>npj Comput. Mater</i>., 7, 185 (2021).<br/>[8] M.H. Rahman et al., <i>APL Machine Learning. </i>2, 016122 (2024).<br/>[9] J. Jang et al., <i>J. Am. Chem. Soc.</i> 142, 18836 (2020).