Data-Driven Discovery of Novel Halide Perovskites for Photovoltaics and Photocatalysis

ABX₃ halide perovskites (HaPs) have been a topic of immense interest over the last two decades by virtue of their diverse optoelectronic applications as well as easy tunability of their electronic properties [1]. HaPs possess a unique combination of attractive properties, showing high absorption coefficients, low carrier recombination rates, high carrier mobility, and a general tolerance to defect states and ion migration [2]. Hybrid organic-inorganic perovskites (HOIPs) are of particular interest, but their long-term stability and performance efficiency remain less than ideal, which motivates comprehensive studies of composition engineering to tailor multiple properties. Such efforts are hindered by the combinatorial nature of the HaP chemical space, given the myriad organic and inorganic cations that may occur at A/B sites, the prospects of cation or anion mixing, and the existence of numerous polymorphs and metastable phases. A computational screening approach that combines high-throughput density functional theory (DFT) and experimental data with machine learning (ML) provides a promising route for success.<br/>In this work, we employed state-of-the-art regression techniques based on Regularized Greedy Forest (RGF) [3] to train composition-based predictive models for the perovskite decomposition energy, band gap, and photovoltaic (PV) efficiency, based on a multi-fidelity dataset of ~ 1300 compounds [3,4] comprising calculations from semi-local and hybrid DFT functionals as well as nearly 100 experimental data points. Inputs to the model include numerical vectors encoding the ABX₃ composition (automatically accounting for mixing at any site), elemental properties of the A, B, and X-site species (with weighted averages used to indicate mixing), and one-hot encoded vectors representing the perovskite phase and data fidelity. In order to extensively span the chemical space, we enumerated a dataset of 151,140 hypothetical HaPs [5] in different phases and used an ensemble of multi-fidelity predictions over 4000 individual models to obtain their properties at a high accuracy. Using these predictions, we performed screening for two applications: (a) suitable absorbers for single-junction solar cells, where low decomposition energy, band gap between 1 and 2 eV, and PV efficiency &gt; 15% were used as the metrics [5], and (b) suitable candidates for photocatalytic water splitting, where the screening criteria were low decomposition energy, suitable band gap and edges (empirically calculated) that straddle the H₂O redox potentials, and high solar-to-hydrogen (STH) efficiency [6]. We performed in-depth DFT computations to validate the best predictions, most of which are mixed cation HOIPs, and arrived at a list of novel compositions with great promise for both applications. The simplicity and accuracy of our models facilitate their application to many different areas of interest, easy expansion to include other ionic species and properties, and ready coupling with targeted experiments to drive rational discovery.<br/> <br/><b>References:</b><br/>[1] J. Yang et al., MRS Bulletin. 47, 940–948 (2022).<br/>[2] Shirayama Masaki, et al. <i>Physical Review Applied</i> 5.1: 014012 (2016).<br/>[3] Johnson Rie et al., <i>IEEE transactions on pattern analysis and machine intelligence</i> 36.5: 942-954 (2013).<br/>[4] J. Yang et al., Digital Discovery. 2, 856-870 (2023).<br/>[5] J. Yang et al., J. Chem. Phys. 160, 064114 (2024).<br/>[6] M. Biswas et al., "Screening of Novel Halide Perovskites for Photocatalytic Water Splitting using Multi-Fidelity Machine Learning", <i>in preparation</i>.