Using automated serendipity to discover how trace water promotes and inhibits lead halide perovskite crystal formation

Philip Nega1, Zhi Li1, Victor Ghosh2,3, Janak Thapa4, Shijing Sun4, Noor Titan Putri Hartono4, Mansoor Ani Najeeb Nellikkal5, Alexander J. Norquist5, Tonio Buonassisi4, Emory M. Chan1, Joshua Schrier3* 1 The Molecular Foundry, Lawrence Berkeley National Laboratory, 1 Cyclotron Road, Berkeley, California, 94720, USA 2 Bronx High School of Science, 75 W 205th Street, The Bronx, New York, 10468, USA 3 Department of Chemistry, Fordham University, 441 E. Fordham Road, The Bronx, New York, 10458, USA 4 Massachusetts Institute of Technology, 77 Massachusetts Avenue, Cambridge, MA, 02139, USA 5 Department of Chemistry, Haverford College, 370 Lancaster Avenue, Haverford, Pennsylvania, 19041, USA


Main text
Halide perovskites are an emerging class of materials 1 of interest for optoelectronics 2 and photovoltaics. 3 Choosing different A-site cations, such as different organoammonium species, yields a wide range of structural and functional diversity beyond the prototypical "perovskite" structure. 4 Halide perovskites are solution processable, which reduces manufacturing costs, and further capital cost reductions could be achieved by reducing the need for environmental controls during device fabrication. 5 However, environmental parameters, such as humidity, remain an important and incompletely understood aspect of perovskite device fabrication. 6 Water affects crystal growth kinetics, morphology, and stability, and is generally considered harmful to device stability and performance, although under certain conditions small amounts of water result in higher quality films and enhanced photovoltaic performances. 5,7,8 This is attributed to water promoting earlier nucleation (which results in low nucleation density and larger grain sizes), whereas anhydrous conditions allow the system to reach a higher degree of supersaturation before nucleation occurs (resulting in increased nucleation site density and smaller grain sizes). 9,10 Post-synthesis water treatment can reorient polycrystalline thin films, improving charge-carrier extraction in photovoltaic devices. 11 However, research has focused primarily on methylammonium and formamidinium lead halides, and so less is known about the two-dimensional (Ruddlesden-Popper and/or Dion-Jacobsen) perovskite materials which have improved long-term photovoltaic stability. 12,13 Single crystal studies provide insight into the growth process, as well as facilitating structure and property determinations for new materials. [14][15][16] Inverse temperature crystallization (ITC) has many advantages for growing large, high-quality single crystals. 17,18 ITC relies on a retrograde solubility effect, in which the product perovskite crystal is less soluble at high temperatures than the precursor species in solution. 19 While previously believed to only occur for methylammonium and formamidinium perovskites, our recent work has found that ITC growth conditions can be used to produce 17 other lead iodide perovskites. 20 Yet, the >1000 articles citing the seminal 2015 papers by Saidmanov et al. 17 and Kadro et al. 18 which introduced ITC growth for halide perovskites have paid scant attention to the role of water. Saidmanov et al.
noted the role of water in perovskite degradation, but did not discuss its role in the ITC process. 17 Kadro et al. noted that the gamma-butyrolactone (GBL) solvent used for ITC is hygroscopic, and cavity nucleation by trace water boiling out during the ITC process might initiate growth. 18 However, using freshly-distilled GBL in an air-and water-free glovebox gave the same results, which was used to exclude this as a primary cause; even if water was present, in this theory it would only promote nucleation. In their mechanistic study of the ITC process, Nayak et al.
attributed the role of trace water to degradation of dimethylformamide (DMF) and GBL solvents into acidic species, which promotes crystal formation by shifting the organoammonium protonation state equilibrium. 19 Their key insight was that this could be achieved more reproducibly by deliberately adding formic acid; again their model implies that water would only promote crystal formation.   Here, we report two counterexamples in which water hinders ITC crystal formation, discovered using a statistical analysis of historical data. Seasonal variations in ambient laboratory conditions influence reaction outcomes, but are generally not controlled and are often only communicated in intra-lab conversations. These variations provide "limited sloppiness" which is the basis for serendipity. 21 Even without the ability to control all possible reaction parameters during each experiment, one can measure and record these data. Given enough observations, statistical methods can be used to identify unexpected correlations.
The combination of high-throughput experimentation and software for comprehensive data capture enables a strategy of data-driven automated serendipity. We previously described our Robot-Accelerated Perovskite Investigation and Discovery (RAPID) system for synthesis of lead iodide perovskites by ITC. 20 Reactions are performed in batches of 96 experiments, typically sampled randomly over the achievable composition space. For the present analysis, reaction outcomes are reduced to a binary score (does a crystal form or not), independent of size or quality. Figure 1a   (e.g., low humidity and high humidity). What is the probability of observing the "wrong" Low humidity crystallization probability High humidity crystallization probability outcome (i.e., fewer total successes in group b than in group a, despite pa>pb )? The probability of k successes for group a is the probability mass function of ( , ! ), denoted ( , , ! ). The probability that as-many-or-fewer-than k successes occur for group b is the cumulative density function of ( , " ), denoted ( , , " ). Both f and F have well-known analytical expressions.
As experiments in each batch are independent, the probability that both events occur is the product, and the probability of more successes in a than in b is By similar logic, the probability that more than k successes are observed for group b is Suppose one is willing to accept a 5% probability of observing the "wrong" outcome (i.e., despite pa < pb, we observe ka > kb), for a given experiment budget of 2n (i.e., n trials assigned to each batch). This demarcation corresponds to setting eqn.  Table I contains a summary of the results and statistical analysis; the electronic Supplementary Materials contains the complete machine-readable description and visualization of the experiment outcomes.
Matched-pair experiments were initially conducted for the dimethylammonium and 4methoxyphenylammonium lead iodide systems by sampling the achievable compositional space uniformly, denoted strategy (a). As indicated in Table I Figure S3). Experiments using strategy (c) revealed differences in crystal formation with and without water for these two systems, with p=0.00026 and p=0.11 for dimethylammonium and 4-methoxyammonium, respectively. Satisfied that at least one of these sampling conditions would highlight the effect of water on crystallization, all three strategies were used for the remaining two systems, acetamidinium and isopropylammonium. In both cases, strategy (b) resulted in the greatest discrepancy between reaction outcomes with added water, (p=0.013 for acetamidinium, and p=0.029 for isopropylammonium).
Different sampling strategies explore regions in compositional space where it is easier or harder to find discordant examples, but the trends should hold over all conditions when the data is combined. Indeed, statistically significant differences in reaction outcome were observed for all four systems studied. Water promotes crystal formation (N+-> N-+) for 4-methoxyammonium and isopropylammonium systems, consistent with expectations from prior ITC results. In contrast, water inhibits crystal formation (N-+ > N+-) for dimethylammonium and acetamidinium systems, contrary to previous expectations. The interventional outcomes agree with the qualitative predictions in Figure 2 in 3 of the 4 tested systems, despite the wildly different sampling strategies employed in the historical dataset. Furthermore, the fewest discrepancies are observed for iso-propylammonium in these interventional experiments, consistent with the difficulty estimate provided by the contour lines in Figure 2. Discrepant reaction outcomes are scattered throughout compositional space in each system (see Figures S4-S16), supporting the first physical interpretation of the statistical model. To explore the relevance for device fabrication, perovskite thin films were fabricated via spin-coating for each of the four cations using the Soltrain system. 32 Separate films were prepared in a glovebox using precursor solutions prepared without water, 1% v/v water, and 2 % v/v water. The same solvents and temperature ranges were used as the ITC counterparts. The grain length distribution of the resulting films was characterized by scanning electron microscopy (SEM). X-Ray diffraction (XRD) patterns indicate that the single-crystal and thinfilm syntheses resulted in the same perovskite phase (see Supporting Materials for relevant experimental details and XRD comparisons). Water has only a slight effect on the grain lengths of 4-methoxyphenylammonium and acetaminidium (see Figure 3). In contrast, water increases grain length in iso-propylammonium perovskite thin films and decreases grain length in dimethylammonium films. The latter is unusual, as most prior studies report water enhancing grain growth, but it is consistent with the corresponding ITC results. In spin-coating experiments, larger grains occur when grain growth is faster than nucleation (typically at lower supersaturation of the perovskite precursors) and smaller grains occur when nucleation is faster than grain growth (typically at higher supersaturation). 33 Water promotes nucleation of isopropylammonium lead iodide at a lower supersaturation concentration, resulting in larger grains.
In contrast, water inhibits nucleation of dimethylammonium lead iodide until the system reaches higher supersaturation, resulting in decreased grain lengths. Despite the different crystallization mechanisms, information obtained about free-standing single crystal formation by ITC can be used to identify systems where additives modify substrate-based grain growth by spin-coating. acetamidinium iodide with 2% water added (i), (j), no water added (m), (n) and isopropylammonium lead iodide with 2% water added (k), (l), no water added (o), (p). Figure S17 is an expanded version of this plot showing results for the 1% water addition.
In summary, water can promote or inhibit perovskite crystallization, depending on the organoammonium cation species present. The latter observation contrasts with previous ITC studies, which attributed the role of water to creation of nucleation sites (from vaporization) or changes in protonation equilibrium (through formation of acid decomposition products), both of which only promote crystal formation. Qualitatively consistent trends in both ITC and thin-film systems suggests a common underlying water-organoammonium interaction mechanism. A practical implication is that trace amounts of water can provide an additional experimental parameter to produce the compact thin films of desired grain sizes needed for stable and efficient devices.
More broadly, this work demonstrates the value of comprehensive electronic experimental records containing both data and metadata. Such records are a prerequisite for automated serendipity, enabling the statistical identification of anomalies that can be subjected to more deliberate study. As such, it serves as additional encouragement for the adoption of automated laboratory processes (such as RAPID 20 and SolTrain 32 ) and software (such as ESCALATE 22 ) that facilitate this type of data collection and reuse.

Supplementary materials
Description of data files and analysis codes; expanded discussion of materials and methods for