Synthesizable Double Perovskite Oxide Search via Machine Learning and High‐Throughput Computational Screening

Abstract

Double perovskite structures have great potential for applications in batteries, lighting devices, and energy‐harvesting materials. In this study, the synthesizability of ABB′O3 double perovskite materials is predicted using machine learning. The machine learning algorithms are validated by performing high‐throughput computational screening. First, material properties extracted from the Materials Project database are used as training data to develop models to predict the formation energy and convex hull energy of general inorganic materials. A regression model predicts the formation energy of general inorganic materials with a high accuracy; an R‐squared value equal to 0.98 and a root‐mean‐square error of 0.175 eV atom−1 are recorded. In addition, a classification accuracy for the convex hull energy of 0.77 is calculated, with an F1‐score of 0.771, in a separate model. Both models are employed to estimate the possible synthesizability of 11 763 ABB′O3 structures and their accuracy is further validated by performing first‐principles calculations, whose classification accuracy for the convex hull energy reaches an accuracy of 0.646, with an F1‐score of 0.733. The constructed surrogate model, as well as the materials database, can guide the discovery of synthesizable double perovskite oxide structures.

Data‐Driven Investigation of the Synthesizability and Bandgap of Double Perovskite Halides

“…For the Hubbard U parameters (GGA + U) treating d electrons in transition metals, the following values were used: [24] 3.32 eV (Co), 3.7 eV (Cr), 5.3 eV (Fe), 3.9 eV (Mn), 4.38 eV (Mo), 6.2 eV (Ni), 3.25 eV (V), and 6.2 eV (W). Although these U values were adopted to be consistent with the MP database and the previous study, [25] further validation was performed by comparing the current bandgap values to those from using U eff values, suggested by Bo Cai et al [36] It demonstrated that the mean absolute difference between two methods was 0.14 eV in Table S3, Supporting Information, confirming that the usage of current U values was adequate. For k-point generation, a gamma-centered Monkhorst-pack was used with grids of 3 × 3 × 3 and 6 × 6 × 6 for geometry optimization and energy calculation, respectively.…”

“…First, 145 chemical descriptors were constructed [ 26 ] using information from the MP database and the DP‐oxide studied earlier. [ 25 ] In addition, the space group number was used to confirm their efficient role in improving the prediction accuracy because chemical descriptors alone cannot identify polymorph structures. Then, the surrogate model for predicting the E form , E hull , and E g of the DP‐halide materials is developed.…”

“…The database constructed for the training set includes 126 328 materials from the MP repository and 11 773 DP‐oxides from our previous work. [ 25 ] For constructing the test set, 13 542 DP‐halide materials among 16 133 double perovskite candidate materials were adopted from a previous study. [ 14 ] Although synthesizable halide materials were proposed by constructing a new tolerance factor, actual validation was not performed.…”

“…Therefore, in our study, we constructed an ML model to find new DP‐halide materials that have not been previously investigated and verified through DFT calculations. Previously built databases (Materials Project [ 24 ] and double perovskite oxide (DP‐oxide) [ 25 ] ) were used as training sets. Then, a prediction model for the thermodynamic stability, formation energy ( E form ), convex hull energy ( E hull ), and bandgap ( E g ) was constructed.…”

Prediction of dielectric constants of ABO3-type perovskites using machine learning and first-principles calculations

“…30 In terms of predicting the synthesizability of perovskites, several approaches have demonstrated that ML with high-throughput computational screening has high functionality and high accuracy. ML can be used in predicting the thermodynamic properties of ABO 3type perovskites, 31,32 classifying the stable and metastable structures of ABO 3 structures based on convex hull energy, 33 and analyzing the correlation between the atomic information and ionic conductivity of ABO 3 perovskites. 34 Also, as a promising candidate for solar light-capturing photovoltaic materials, 11 Li(Na)-based perovskites among 15 000 candidates were selected through classification on the formation energy and bandgap.…”