Perovskite synthesizability using graph neural networks

Abstract

Perovskite is an important material type in geophysics and for technologically important applications. However, the number of synthetic perovskites remains relatively small. To accelerate the high-throughput discovery of perovskites, we propose a graph neural network model to assess their synthesizability. Our trained model shows a promising 0.957 out-of-sample true positive rate, significantly improving over empirical rule-based methods. Further validation is established by demonstrating that a significant portion of the virtual crystals that are predicted to be synthesizable have already been indeed synthesized in literature, and those with the lowest synthesizability scores have not been reported. While previous empirical strategies are mainly applicable to metal oxides, our model is general and capable of predicting the synthesizability across all classes of perovskites, including chalcogenide, halide, and hydride perovskites, as well as anti-perovskites. We apply the method to identify synthesizable perovskite candidates for two potential applications, the Li-rich ion conductors and metal halide optical materials that can be tested experimentally.

Challenges, Opportunities, and Prospects in Metal Halide Perovskites from Theoretical and Machine Learning Perspectives

“…To overcome this difficulty, a recent study suggested that an ML model based on transfer learning is able to predict the synthesizability of perovskites with desirable band gaps. [ 52 ]…”

“…Charge-, spin-dependent ML potentials, [45][46][47][48] Ferroic phases Long-range charge/dipole interactions Charge-dependent ML potential [45,46] Excited states dynamics Hot carrier cooling Accurate description of non-adiabatic process, Nanoscale NAMD Excited states ML potential [49] PL mechanism of 0D MHPs Nanodomains Spin-dependent excited states ML potential [48,49] Materials design Holistic ML prediction High dimensional chemical space, Multiple property predictions Geometric learning, [50] Multi-task [51] /Transfer learning [52] High quality dataset for MHPs Data mining of experimental results, Dataset at high levels of theory Natural Language Processing, [53,54] Distributed data mining [55] Trial-and-error synthesis Controllable synthesis in various conditions, Automatic optimization of solution-based synthesis ML optimization for robotics synthesis [56][57][58] Adv. Energy Mater.…”

“…The early applications of ML in chemistry and materials science were to build a supervised regression ML model to predict a target chemical property, such as transmittance, [32] synthesizability, [52] and band gap. [80,81] Transfer learning is one promising ML approach to training an ML model with a small number of data set where general knowledge is transferred from the base model which has a large data set (Figure 1b).…”

Free energy predictions for crystal stability and synthesisability

“…There has been successes in using information from failed experiments, 89 but probabilistic models are more generally built from positive and unlabelled data for specific classes of material. 90,91 For more universal models, a large amount of diverse data is required. There has been progress in text mining of synthesis information from the literature.…”

Molecular screening for solid–solid phase transition by machine learning

“…In this case, positive-unlabeled (PU) classification is a reasonable problem setting affording a discriminant boundary between positive and unlabeled data 24 . We can determine the possibility of solid-solid phase transition in unlabeled data in a manner similar to using the PU setting to predict the synthesizability of materials 25,26 . We used four different models to decide on the discriminant boundary: random forest (RF), neural network (NN), support vector machine (SVM), and gradient boosting decision tree (GBDT) (Table S1).…”