Metal halide perovskites are a promising class of materials for next-generation photovoltaic and optoelectronic devices. The discovery and full characterization of new perovskite-derived materials are limited by the difficulty of growing high quality crystals needed for single-crystal Xray diffraction studies. We present the first automated, high-throughput approach for metal halide perovskite single crystal discovery based on inverse temperature crystallization (ITC) as a means to rapidly identify and optimize synthesis conditions for the formation of high quality single crystals. Using this automated approach, a total of 1928 metal halide perovskite synthesis reactions were conducted using six organic ammonium cations (methylammonium, ethylammonium, nbutylammonium, formamidinium, guanidinium, and acetamidinium), increasing the number of metal halide perovskite materials accessible by ITC syntheses by three and resulting in the formation of a new phase, [C2H7N2][PbI3]. This comprehensive dataset allows for a statistical quantification of the total experimental space and of the likelihood of large single crystal formation. Moreover, this dataset enables the construction and evaluation of machine learning models for predicting crystal formation conditions. This work is a proof-of-concept that combining high throughput experimentation and machine learning accelerates and enhances the study of metal halide perovskite crystallization. This approach is designed to be generalizable to different synthetic routes for the acceleration of materials discovery.

Autonomous experimentation systems use algorithms and data from prior experiments to select and perform new experiments in order to meet a specified objective. In most experimental chemistry situations, there is a limited set of prior historical data available, and acquiring new data may be expensive and time consuming, which places constraints on machine learning methods. Active learning methods prioritize new experiment selection by using machine learning model uncertainty and predicted outcomes. Meta-learning methods attempt to construct models that can learn quickly with a limited set of data for a new task. In this paper, we applied the model-agnostic meta-learning (MAML) model and the Probabilistic LATent model for Incorporating Priors and Uncertainty in few-Shot learning (PLATIPUS) approach, which extends MAML to active learning, to the problem of halide perovskite growth by inverse temperature crystallization. Using a dataset of 1870 reactions conducted using 19 different organoammonium lead iodide systems, we determined the optimal strategies for incorporating historical data into active and meta-learning models to predict reaction compositions that result in crystals. We then evaluated the best three algorithms (PLATIPUS and active-learning k-nearest neighbor and decision tree algorithms) with four new chemical systems in experimental laboratory tests. With a fixed budget of 20 experiments, PLATIPUS makes superior predictions of reaction outcomes compared to other active-learning algorithms and a random baseline.

Halide perovskite materials have attracted great interest for applications in low-cost, solution-processed solar cells and other optoelectronics applications. The role of moisture in perovskite device degradation and crystal formation processes remains poorly understood.Here we use a data-driven approach to discover the influence of trace amounts of water on perovskite crystal formation by analyzing a comprehensive dataset of 8,470 inversetemperature crystallization lead iodide perovskite synthesis reactions, performed over 20 months using a robotic system. We identified discrepancies between the empirical crystal formation rate in batches of experiments conducted under different ambient relative humidity conditions for each organoammonium cation. We prioritized these using a statistical model, and then used the robotic system to conduct 1,296 controlled interventional experiments in which small amounts of water were deliberately introduced to the reactions. The addition of trace amounts of water promotes crystal formation for 4-methoxyphenylammonium lead iodide and iso-propylammonium lead iodide and inhibits crystal formation for dimethylammonium lead iodide and acetamidinium lead iodide. We also performed thinfilm syntheses of these four materials and determined the grain size distributions using scanning electron microscopy. The addition of water results in smaller grain sizes for dimethylammonium and larger grain sizes for isopropylammonium, consistent with earlier or delayed nucleation, respectively. The agreement between the inverse temperature crystallization and thin film results indicates that this is a feature of the ammonium-water interaction that persists despite differences in the synthesis method.

Additives in the precursor solution can promote lead-halide perovskite (LHP) crystallization. We present a systematic exploration of nine (9) bipyridine- and terpyridine-based additives selected from 29 candidates using high-throughput single-crystal growth. To combat selection bias and generate hypotheses for future experimental cycles of learning, we featurize candidate additives using Mordred descriptors and compare similarity metrics. A previously unreported additive, 6,6′-dimethyl-2,2′-dipyridyl, is shown to work particularly well (the highest top 10th percentile is ∼3.8 mm, in comparison to ∼1.9 mm without additive) in improving the crystallization of prototypical methylammonium lead iodide (MAPbI3). Our strategy of machine-learning-guided high-throughput experimentation is generally applicable to other crystal growth problems.

Machine learning is a useful tool for accelerating materials discovery, however it is a challenge to develop accurate methods that successfully transfer between domains while also broadening the scope of reaction conditions considered. In this paper, we consider how active- and transfer-learning methods can be used as building blocks for predicting reaction outcomes of metal halide perovskite synthesis. We then introduce a serendipity-based recommendation system that guides these methods to balance novelty and accuracy. The model-agnostic recommendation system is tested across active- and transfer-learning algorithms, using laboratory experiments for training and testing and a time-separated hold out that includes four different chemical systems. The serendipity recommendation system achieves high accuracy while increasing the scope of the synthesis conditions explored.