Machine Learning-Directed Navigation of Synthetic Design Space: A

Subscriber access provided by The University of Melbourne Libraries

Article

Machine-Learning-Directed Navigation of Synthetic Design Space: A Statistical Learning Approach to Controlling the Synthesis of Perovskite Halide Nanoplatelets in the Quantum-Confined Regime Erick J. Braham, Junsang Cho, Kristel M. Forlano, David F. Watson, Raymundo Arroyave, and Sarbajit Banerjee Chem. Mater., Just Accepted Manuscript • DOI: 10.1021/acs.chemmater.9b00212 • Publication Date (Web): 11 Apr 2019 Downloaded from http://pubs.acs.org on April 11, 2019

Just Accepted “Just Accepted” manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides “Just Accepted” as a service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. “Just Accepted” manuscripts appear in full in PDF format accompanied by an HTML abstract. “Just Accepted” manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are citable by the Digital Object Identifier (DOI®). “Just Accepted” is an optional service offered to authors. Therefore, the “Just Accepted” Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the “Just Accepted” Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these “Just Accepted” manuscripts.

is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

Page 1 of 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

Chemistry of Materials

Machine-Learning-Directed Navigation of Synthetic Design Space: A Statistical Learning Approach to Controlling the Synthesis of Perovskite Halide Nanoplatelets in the Quantum-Confined Regime Erick J. Braham,†,§# Junsang Cho,†,§# Kristel M. Forlano,† David F. Watson,|| Raymundo Arròyave,§ Sarbajit Banerjee†,§* †Department

of Chemistry, Texas A&M University, College Station, TX, 77843-3255, USA of Materials Science and Engineering, Texas A&M University, College Station, TX 77843-3255, USA ||Department of Chemistry, University at Buffalo, The State University of New York, Buffalo, New York 14260-3000, USA §Department

#these

authors contributed equally to this work

ABSTRACT: The design of a chemical synthesis often relies on a combination of chemical intuition and Edisonian trial-and-error methods. Such methods are not just inefficient but inherently limited in their ability to quantitatively predict synthetic outcomes, easily defeated by complex interplays between variables, and oftentimes based on suppositions that are limited in validity. The synthesis of nanomaterials has been especially prone to empiricism given the combination of complex chemical reactivity as well as mesoscopic nucleation and growth phenomena spanning multiple temporal and spatial dimensions. Here, utilizing the synthesis of 2D CsPbBr3 nanoplatelets as a model system, we demonstrate an efficient machine learning navigation of reaction space that allows for predictive control of layer thickness down to sub-monolayer dimensions. Support vector machine (SVM) classification and regression models are used to initially separate regions of the design space that yield quantum confined nanoplatelets from regions yielding bulk particles and subsequently to predict the thickness of quantum confined CsPbBr3 nanoplatelets that can be accessed under specific reaction conditions. The SVM models are not just predictive and efficient in sampling the available design space but also provide fundamental insight into the influence of molecular ligands in constraining the dimensions of nanocrystals. The results illustrate a quantitative approach for efficient navigation of reaction design space and pave the way to navigation of more elaborate landscapes beyond dimensional control spanning polymorphs, compositional variants, and surface chemistry.

INTRODUCTION Scaling periodic solids to nanometer-sized dimensions gives rise to distinctive quantum confinement effects, an increased proportion of atoms residing at surfaces, and the elimination of extended defects. Such phenomena have spurred intense interest in the programmable growth and assembly of nanomaterials with potential applications in optoelectronics, medical diagnostics, catalysis, and energy harvesting. Colloidal nanocrystal synthesis represents an important tool in the arsenal of synthetic materials chemists and provides access to welldefined monodisperse nanocrystals with surfaces passivated by coordinating ligands or electrostatically bound surfactants.1–4 The synthesis of nanocrystals from molecular precursors is generally understood according to variations of nucleation and growth theory with the added ligands enabling the temporal separation of the two steps, stabilizing specific crystallographic facets through preferential binding, and providing control over monomer supersaturation.5–9 Achieving precise synthetic control over the shape and size distributions of these ensembles of nanocrystals is imperative for the effective utilization of such materials. Recent efforts have sought to expand mechanistic understanding and establish correlations between precursor reactivity and the eventual crystal structure, dimensionality, shape, and surface chemistry of nanocrystals; nevertheless,

much of synthetic nanochemistry remains strongly reliant on the development of empirical synthetic strategies.10–14 Edisonian trial-and-error methods involving changing a single synthetic variable and observing the response are standard practice but represent a rather inefficient means of exploring potentially vast design spaces. Such methods are furthermore limited in their ability to quantitatively predict synthetic outcomes and do not provide a satisfactory understanding of variable correlations, factors underpinning challenges with reproducibility, and parameters necessary to facilitate the application of modern process design tools. Some initial attempts at bringing statistical learning to nanocrystal synthesis have invoked design of experiments (DOE) methods such as full and fractional factorial sampling coupled with ridge regression to establish correlations between synthetic variables and nanocrystal dimensions.15,16 These methods provide a means of rapidly exploring synthetic correlations but impose specific sampling constraints on creation of the model and often interpolate large areas of the design space. In contrast, machine learning approaches are more versatile in enabling the use of incomplete (and sparse) datasets not acquired according to specific constraints allowing for the creation of robust models for the prediction of quantitative or qualitative synthetic outcomes. In this article, we demonstrate the application of a

1 Environment ACS Paragon Plus

Chemistry of Materials 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

non-linear data-driven machine learning model to predictively control the layer thickness of perovskite halide nanoplatelets. The intersection of machine learning and chemistry has provided new opportunities spanning the range from atomistic design of solid-state compounds to elucidation of ligand parameters underpinning molecular catalysis17–19 and mapping of compositional gradients in hyperspectral maps of discharged cathode materials.20 Nonlinear modeling, such as using support vector machines for classification or regression, has proven to be successful at predicting a wide range of materials properties based on elemental parameters and atomic interactions ranging from crystal structure to band gap.21–25 However, such methods have rarely been applied to a singular chemical synthesis due to the cost of creating a viable dataset large enough to both provide new chemical insights and avoid overfitting. By utilizing existing synthetic results as initial datasets, input data can be efficiently acquired to allow for the meaningful use of robust nonlinear modeling techniques. Size control in nanocrystal synthesis provides an excellent output for evaluating quantitative predictive models developed based on existing data. The synthesis of 2D CsPbBr3 perovskite nanoplatelets is used here as a model system given the clear layer-thicknessdependent photoluminescence spectral signatures that allow for rapid evaluation of size distributions within samples.26–28 Lead halide perovskites have attracted extensive interest as a result of their remarkable properties including tunable exciton binding energies, high oscillator strengths of bandgap transitions, narrow emission bands, and high photoluminescence quantum yields.29–34 CsPbBr3 is a stable allinorganic perovskite accessible from solution-phase synthesis; dimensional confinement below the characteristic Bohr radius of 3.5 nm brings this material to the quantum confined regime and allows for tunability of the bandgap as a function of the layer thickness.35 Discrete photoluminescence emission bands characteristic of each layer thickness indicate a pronounced modulation of the bandgap from 2.36—3.20 eV upon scaling from the bulk material down to monolayer platelets.36,37 Control of the layer thickness is generally accomplished through ligandassisted synthesis.35,38,39 In this work, utilizing photoluminescence spectroscopy as the primary probe, we demonstrate a predictive, statistically informed model for synthetic control of particle thickness developed using a sparse dataset. A machine learning approach for exploration of multivariate design space further provides mechanistic insight into the role of the ligand shell in inducing dimensional confinement.

RESULTS AND DISCUSSION Mapping Multivariate Reaction Space Figure 1A represents an initial data set for the synthesis of CsPbBr3 nanoplatelets prepared using a hot colloidal method where multivariate reaction parameters have been explored by independently varying the reaction temperature (50, 100, and 150°C at a fixed Pb:alkylamine ratio of 1:20), ligand concentration (in the 1:x range spanning x = 5—40 (wherein x is concentration of alkylamine) at a constant temperature of 100°C), and chain-lengths of n-alkylamine ligands (C4, C8, C12, and C18). The addition of coordinating amine ligands brings about pronounced dimensional confinement as schematically illustrated in Figure 1B.38,45 Colloidal CsPbBr3 nanoplatelets adopts a distorted orthorhombic crystal structure upon dimensional confinement, which stands in contrast to the cubic phase stabilized at high

Page 2 of 15

temperature (>130°C) and tetragonal phase at intermediate temperatures (88-130°C) of bulk CsPbBr3.46 The powder X-ray diffraction patterns (XRD) in Figure 1C for CsPbBr3 nanoplatelets grown using alkylamines with different chain lengths can be indexed to the orthorhombic phase.38,46,47 The average layer thickness (navg.) of 2D CsPbBr3 nanoplatelets is expressed in terms of the number of [PbBr6]4- octahedral layers. Values of navg. plotted in Figure 1A been determined from the photoluminescence (PL) emission spectra as described in the experimental section and have been further verified by electron microscopy.38 The photoluminescence emission spectra have been fit to multiple Gaussian distributions centered on the wavelengths of the discrete emission bands to deconvolute the fractional contribution of individual emission bands. The correlation between navg. and the thickness determined by TEM (tm) is illustrated for five representative samples in Figures 1D—M. Figures 1D—H exhibit the photoluminescence spectra and Figures 1I—M show the corresponding TEM images. Unlike II-VI quantum dots, the size distribution of CsPbBr3 2D nanoplatelets is not continuous but instead discretized based on the layer thicknesses. Emission spectra acquired at each set of reaction conditions are characterized by multiple PL emission bands quantized to populations of different layer thicknesses and manually assigned to peak positions as previously determined in single-particle PL emission spectra measurements.28,36,38 Figure S1 plots the PL emission wavelength of 2D CsPbBr3 nanoplatelets measured from singleparticle measurements versus the vertical dimension of nanoplatelets.38 In the strongly quantum confined regime (below n = 10 and 6 nm), the energy bandgap (and consequently the photoluminescence resulting from bandgap emission) varies strongly as a function of layer thickness. As the vertical dimension of CsPbBr3 nanoplatelets increases to greater than the exciton Bohr diameter, the PL emission converges to a maximum wavelength of approximately 525 nm (which reflects the 2.36 eV bandgap of bulk CsPbBr3) and no longer varies with thickness. Emission bands at 525 nm are assigned to n = 10 layers when computing navg; the validity of this assignment is further examined below. It is worth noting that the type of alkylamine ligand induces subtle shifts of PL emission maxima owing to (i) variations in types and concentrations of trap states, (ii) differences in dielectric constants and surface dipoles of the ligand shell, (iii) extent of electronic coupling between nanoplatelets.38,48,49 Instead of fitting to the same peak maximum across all conditions, each spectrum has been manually fitted to account for these shifts of the PL emission maxima. Figure S2 shows the fitting residual obtained from subtracting the multi-Gaussian fit sum from the experimental PL emission spectrum for the CsPbBr3 nanoplatelets synthesized using C12 with 1:40 at 100 °C demonstrating that the residuals are negligible and distributed randomly around zero. This procedure (in contrast to unconstrained automated fitting protocols) yields discrete bands correspondent with the single-particle emission spectra. The oscillator strength of 2D CsPbBr3 nanoplatelets can vary somewhat as a function of layer thickness depending on the specific synthetic route. However, there is considerable literature precedent of the PL quantum yield (QY) of CsPbBr3 perovskite nanocrystals in the quantum confined regime being within the same order of magnitude (33 % for n = 3, 44.7 % for n = 4, 31 % for n = 5) since these nanocrystals are not as prone to surface traps as II-VI semiconductors. Indeed, the elimination of trap states and the resulting defect-tolerant band

2

ACS Paragon Plus Environment

Page 3 of 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


structure is pivotal to the utilization of the relative intensities of discrete PL bands as a reliable proxy for the fractional contributions of different layer thicknesses.28,37,40,50,51 Figure S3 exhibits additional TEM images acquired for the five samples noted in Figure 1. The average layer thicknesses from left to right are navg. = 2.8 (tm = 1.9±0.4 nm), navg. = 3.5 (tm = 2.3±0.4 nm), navg. = 3.5 (tm = 2.5±0.3 nm), navg. = 5.1 (tm = 3.0±0.3 nm), and navg. = 7.9 (tm = 10.3±0.7 nm) ordered from thinner to thicker nanoplatelets. A strong correlation is observed between the spectroscopic measurements and TEM imaging of layer thicknesses when the dimension of 2D nanoplatelets is below the exciton Bohr diameter in the strongly quantum confined regime, which corroborates the validity of the former method in providing a quantitative metric of layer thickness in the quantum confined regime. Table S1 summarizes the results of the syntheses, shown in Figure 1A, and the estimated populations corresponding to the different layer thicknesses. Figure 1 illustrates the clear influence that the choice of the ligand exerts on the extent of

dimensional confinement of 2D CsPbBr3 nanoplatelets. In general, higher ligand concentrations, longer alkyl chain lengths, and lower reaction temperatures are observed to result in the stabilization of few-layered nanoplatelets in the strongly quantum confined regime. In contrast, lower ligand concentrations, shorter chain lengths, and higher reaction temperatures bring about rapid growth of thicker nanoplatelets approaching the bulk limit. The extent of dimensional confinement is generally understood to depend on the crystalline order of the self-assembled monolayer of ligand molecules and their dynamic equilibrium with free ligand molecules in solution.45 The structure of the ligand shell, the extent of monomer supersaturation, and the rate of monomer addition are in turn determined by the length of the alkyl chains, their branching, the reaction temperature, and concentrations. In this article, we have sought to develop a predictive machine learning model that captures the complex interplay between these reaction parameters.

Figure 1. Sampling of the multivariate design space for the synthesis of CsPbBr3 nanoplatelets. (A) 3D visualization of the average layer thickness as a function of the ligand chain-length and concentration at different reaction temperatures with added contrast on overlapping data points; (B) schematic illustration of ligand-induced dimensional confinement; (C) XRD patterns of CsPbBr3 nanoplatelets as a function of the alkylamine chain length; the orange ticks on the horizontal axis correspond to the reflections of orthorhombic CsPbBr3 with PDF# 01−072−7929. (D)-(H) PL emission spectra of CsPbBr3 nanoplatelets grown under different reaction conditions and (I)-(M) corresponding TEM images obtained for these samples. The photoluminescence emission spectra have been obtained at an excitation wavelength of 360 nm and are fitted to multi-peak Gaussian lineshapes representing the characteristic emission bands of different layer thicknesses.

3



In addition to the vectoral exploration shown in Figure 1A, 34 additional samples were synthesized randomly expanding the dataset to areas of design space that do not fall along the initial exploration vectors and are summarized in Table S2. These selections were not selected for design space optimization but in the hope that a more random dataset would allow for better sampling of the design space and improve the robustness of the machine learning modeling methods. Classification, Curation, and Feature Set Development The initial dataset of 74 samples has been used to build two models of the reaction design space: (a) a classifier to define a boundary between bulk and quantum confined particles and (b) a regressor to predict the thickness of particles in the quantum confined regime. Figure 2 depicts the steps involved in the modeling procedure beginning with curation of the dataset and feature set development, followed by modeling of both the classifier and regressor, and finally, validation of the regressor using samples randomly selected from the quantum confined regime as determined by the classifier.

Page 4 of 15

ternary Cs-Pb-Br system such as stabilization of the 0D Cs4PbBr6 phase. Clear trends are discernible in Figure 1A for the independent variables, ligand choice, ligand concentration, and temperature, in terms of their ability to influence the particle size.38 A metric needs to be defined that best captures the influence of the ligand chemical structure and composition on the kinetics of crystal growth to allow for regression and interpolation to new values. A series of values have been tabulated for the n-alkylamine physical and chemical properties (Table S3) and compared using both a t-test, as shown in Table 1, and by error analysis of a models calculated with each descriptor (a wrapper), the results of which are presented in Tables S4 and S5. The classes being compared in the t-test have been created by curating the dataset into two sections: a rigorously quantum confined class comprising samples where 5 have large underestimation errors. The SVM regressor (Fig. 4D) performs the best with a RMSE-CV of 0.7556 layers but still shows some underestimation of layer thicknesses for navg. > 5 layers. This regression is the top performer of all the tested machine learning models shown in Tables S4 and S5. This lead model has an aforementioned RMSE-CV of 0.7556 layers, a coefficient of determination of 0.895, a gamma of 0.5, a cost of 2.1 and a support vector count of 48.

ACS Paragon Plus Environment 5


Page 6 of 15

Figure 4. Leave-one-out cross validation results for four different regression models predicting the synthesis of CsPbBr3: (A) linear regression (R2 = 0.645), (B) quadratic regression (R2 = 0.838), (C) quadratic regression with cross terms (R2 = 0.861), and (D) support vector machine regression (R2 = 0.965). The Y = X line is delineated as a representation of an ideal prediction. R2 values are for the model fits to the experimental data with the full 57 data point set.

The difficulties in accurate prediction of thicker layers derives from assumptions made in inclusion of bulk photoluminescence emission as a constant 10 layers, which thereby introduces greater error in samples with proportionately larger bulk contributions. Note that this is essentially a limitation of the spectroscopic method, which allows for clear delineation of layer thicknesses near the quantum confined regime but does not allow for multilayered platelets to be distinguished in the bulk regime. Figure S4 plots the RMSE leave-one-out cross validation as a function of the estimated bulk thickness demonstrating that the estimate of 10 layers is the best option for a bulk estimation to avoid skewing the data to higher averages (which would increase the cross-validation error). Despite this underestimation, the RMSE values of the cross-validation analysis indicate that the SVM model has achieved a reasonable level of predictivity for this dataset achieving sub-monolayer accuracy. Video S1 exhibits the four-dimensional design space mapped by the SVM model indicating the evolution of the layer thickness as a function of the diffusion coefficient and ligand concentration with increasing temperature. Figure 5 plots contour SVM regressor slices from this model. Examination of the SVM model built from the 57 quantum confined samples indicates that the model approximation of the response space is influenced by each independent variable. Figure 5A shows a slice of the regressor output at 50°C with the color map corresponding to the predicted layer thickness for the given synthetic conditions. A minimum of 3.0 average layers can be observed at an approximate ligand concentration of Pb:alkylamine of 1:30 for a ligand diffusion coefficient of 0.85×10-10 m2/s. The location of this minimum indicates that at 50°C, increasing the concentration and the chain length

(decreasing diffusion coefficient) yields smaller predicted thicknesses (up until a precursor:ligand ratio of 1:30 and diffusion coefficient of 0.85×10-10 m2/s, respectively). Further increases in ligand concentration and chain length bring about an increase in thickness. The increase observed for 1:x concentrations greater than 1:30 is most likely a result of (i) the rapid formation of the ligand shell at high ligand concentrations, which results in shells with less regular packing that are more pervious to monomer addition, as well as (ii) the high concentration of complexed monomeric species. Both these factors result in more rapid crystal growth yielding thicker nanoplatelets. The increase in thickness with diffusion coefficients lower than 0.85×10-10 m2/s, corresponding to higher chain lengths, is furthermore a reflection of the increased disorder of ligand shells for longer-chain alkylamines at low temperatures. Disordered ligand shells with lower packing densities allow for easier monomer addition, which results in the growth of thicker nanoplatelets (Fig. 6).38 Indeed, the selfassembly of n-alkylamines on surfaces proceeds through initial adsorption and desorption steps until the system approaches quasi-equilibrium close-packed conditions as sketched in Figure 6 wherein dispersive interactions have been maximized along the chain lengths, thereby yielding the most optimal enthalpic stabilization to offset entropic losses from conformational restrictions.52 For longer chain amines at lower temperatures, thermal desorption is more hindered and thus kinetically trapped imperfectly ordered ligand shells are more likely to be stabilized, allowing for an increased rate of monomer addition and faster crystal growth.38,52,53 Remarkably, the SVM model accurately captures this complex interplay between enthalpic and entropic factors.


6

Page 7 of 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


Figure 5. Contour plot slices of the SVM regression at temperatures of (A) 50°C, (B) 82°C, (C) 120°C, and (D) 150°C. Video S1 shows the continuous evolution of the contours as a function of temperature. = 40 induces a phase transformation from CsPbBr3 to an The correlation of average layer thickness of CsPbBr3 altogether different lead-deficient Cs4PbBr6 structure as a result nanoplatelets to synthetic conditions is not intuitive given the of the amine ligands mediating leaching of surficial PbBr2 trade-offs between entropic and enthalpic factors, kinetics and thermodynamics of ligand shell assembly (Fig. 6J), and layers.41,54 Additionally, the longest chain length ligands (e.g., sensitivity of the stabilized phase to ligand denticity, C14 or C18) yield thicker nanoplatelets at low reaction concentration, and branching .45 In a similar vein, temperature temperatures in the range of 50—100°C since at low both accelerates growth kinetics and monomer diffusion while temperatures, a kinetically trapped disordered ligand shell is allowing for stabilization of more ordered ligand shells, which stabilized (and cannot be rapidly equilibrated to a retards monomer addition. Naively, the highest concentration thermodynamically stable self-assembled monolayer), which and longest chain length of ligands can be anticipated to allows for facile diffusion of monomeric species.38 At higher stabilize the thinnest nanoplatelets by effectively buffering temperatures, longer-chain alkylamines are able to form monomer supersaturation, regulating monomer diffusion, and ordered ligand shells as a result of thermally facilitated constituting an extended ligand shell. However, the highest desorption,52,53 allowing for quasi-equilibrium conditions to be ligand concentrations or longest chain ligands do not always reached. However, monomer diffusion is also greatly enhanced yield the thinnest nanoplatelets. Figure 5B shows model results as expected for temperature-variant Fickian diffusion.55 The at 82°C where the global minimum of the model of 2.4 average local minima reflects the achievable balance between these two layers is observed for similar ligand concentrations and competing influences representative of thermodynamic diffusion coefficients as in Figure 5A indicating that increasing equilibrium and kinetic metastable regimes. In other words, the temperature up to 82°C allows for the most effective correlations provided by the regression model allow for confinement of crystal growth. Figures 5C and D plot slice qualitative mechanistic understanding to be gleaned from the contour maps at higher temperatures of 120 and 150°C, SVM model in a non-intuitive manner. respectively; the predicted layer thickness increases with To illustrate some of the mechanistic regimes, Figure 6 temperature across the entire response space in this range. provides a schematic of the 82°C and 150°C slices of the SVM The stabilization of relatively thicker layers results from contour plot also shown in Figures 5B and 5D respectively. increased diffusion coefficients of monomeric species, which These illustrations show the entropic effects of ligand facilitates faster crystal growth. The ligand concentration and concentration and chain length at the two different chain length are correlated in the temperature range between temperatures. Figure 6J shows a general schematic of the 76—150°C with the local minima shifting to lower nucleation and growth process differentiating the stabilization concentrations and longer chain lengths with increasing of ordered ligand shells or kinetically trapped disordered states. temperature, which is reflective of the interplay between Figures 6A-I schematically illustrate different ligand shell formation of a well-ordered ligand shell and the mobility of configurations and their ability to modulate monomer addition. monomeric species. The highest alkylamine concentration of x



Page 8 of 15

Figure 6: Schematic of crystal growth regimes corresponding to the 82 and 150°C slices of the RBF-SVM model. Illustrations of different regimes are shown as follows (darker red monomers imply faster diffusion as a result of higher temperatures): (A) high ligand concentration results in a disordered monolayer; (B) conditions allowing for optimal ligand shell monolayer formation resulting in the global minimum of layer thicknesses accessed in this work; (C) high temperature and concentration as well as longer ligand chain lengths yield disordered layers whilst facilitating relatively rapid monomer diffusion; (D) longer chain length increases packing disorder (corresponding to stabilization of a kinetic product); (E) low ligand concentration enables facile crystal growth at sub-monolayer coverages; (F) short chain length allows desorption and passivation as well as easier monomer transport, thereby favoring crystal growth; (G) at high temperatures, long ligand chain lengths allows for stabilization of somewhat ordered ligand shells thereby limiting crystal growth; (H) low ligand concentration and high temperature favors ligand desorption enabling facile monomer addition and crystal growth; (I) high temperature and short ligand chain length favors facile monomer addition and enables rapid crystal growth; (J) schematic illustration of ligand shell formation alternatively yielding an ordered monolayer maximizing dispersive interactions (thermodynamic product) or becoming trapped within a disordered state representing the kinetic product. of the limitations of the spectroscopic method delineated above. In order to test for overfitting and analyze the predictivity of The average layer thicknesses (navg.) of 2D perovskite the regressors, a validation test set of six newly synthesized nanoplatelets in the validation set as deduced from ensemble PL samples that had no influence on the creation of the models has emission spectra and determined from TEM imaging (t) in been selected and analyzed. The six samples have been selected Figure 7 are listed in Table S6. Figure S5 shows additional to be in the 6, corresponding to particles above the quantum confined regime, a layer thickness of n = 10 has been approximated. Assignments of layer thicknesses (ni) have been made based on previously reported single-particle photoluminescence maximum emission wavelengths.38 The integrated area of the photoluminescence emission band corresponding to a specific layer thickness (ni) is divided by the total integrated area of the photoluminescence spectrum to yield the relative proportion of each layer (ni) within the sample. Tables S1 and S2 list the deconvoluted areal intensities for the different syntheses examined in this work. Gaussian lineshapes have been used to fit emission bands in each case. All statistical models were computed using R 3.4.1 with linear (eq. 4) and quadratic (eqs. 5-6) regressions of the following form computed using a linear modeling function to determine the coefficients of the regressions: 𝑦 = 𝑏1 + 𝑏2𝑥1 + 𝑏3𝑥2 + 𝑏4𝑥3 … (4)

contour plot slices for temperatures spanning 50—150°c of the SVM regression model can be found in the Supporting Information.

AUTHOR INFORMATION Corresponding Author * [email protected]

Author Contributions #These authors contributed equally.

ACKNOWLEDGMENTS We gratefully acknowledge support from the National Science Foundation under DMREF 1627197. EJB, RA, and SB acknowledge support from the Data-Enabled Discovery and Design of Energy Materials (D3EM) program funded by the National Science Foundation under DGE-1545403.

REFERENCES (1) Murray, C. B.; Kagan, C. R.; Bawendi, M. G. Synthesis and Characterization of Monodisperse Nanocrystals and Close-Packed Nanocrystal Assemblies. Annu. Rev. Mater. Sci. 2000, 30, 545– 610. (2) Jana, N. R.; Gearheart, L.; Murphy, C. J. Seed𝑦 = 𝑏1 + 𝑏2𝑥1 + 𝑏3𝑥21 + 𝑏4𝑥2 + 𝑏5𝑥22 + 𝑏6𝑥3 + 𝑏7𝑥23 … (5) Mediated Growth Approach for Shape-Controlled Synthesis of Spheroidal and Rod-like Gold 2 2 2 Nanoparticles Using a Surfactant Template. Adv. 𝑦 = 𝑏1 + 𝑏2𝑥1 + 𝑏3𝑥1 + 𝑏4𝑥2 + 𝑏5𝑥2 + 𝑏6𝑥3 + 𝑏7𝑥3 + 𝑏8𝑥1 Mater. 2001, 13, 1389–1393. 𝑥2 + 𝑏9𝑥1𝑥3 + 𝑏10𝑥2𝑥3 … (6) (3) Yin, Y.; Alivisatos, A. P. Colloidal Nanocrystal The variables xn represent the three input variables and bn Synthesis and the Organic-Inorganic Interface. represent the coefficients of the regression. Eq. 5 includes Nature 2005, 437, 664–670. calculation of coefficients b8-10 allowing for the inclusion of (4) Stowell, C. A.; Korgel, B. A. Iridium Nanocrystal variable correlation. The T-test was performed assuming equal Synthesis and Surface Coating-Dependent Catalytic variance as a standard two sample t-test (eqs. 7-8). Activity. Nano Lett. 2005, 5, 1203–1207. 𝑋1 ― 𝑋2 (5) Kort, K. R.; Banerjee, S. Ligand-Mediated Control 𝑇= … (7) 1 1 𝑠𝑝 𝑛 + 𝑛 of Dislocation Dynamics and Resulting Particle 1 2 Morphology of GdOCl Nanocrystals. Small 2015, 11, 329–334. 2 2 (𝑛1 ― 1)𝑠𝑋1 + (𝑛2 ― 1)𝑠𝑋2 (6) Yu, W. W.; Wang, Y. A.; Peng, X. Formation and 𝑠𝑝 = … (8) 𝑛1 + 𝑛2 ― 2 Stability of Size-, Shape-, and Structure-Controlled The variable X̄ represents the sample means, n the sample sizes CdTe Nanocrystals: Ligand Effects on Monomers and s the sample standard deviations with sp the pooled and Nanocrystals. Chem. Mater. 2003, 15, 4300– variance. The e1071 R package version 1.6-8 was utilized to 4308. compute the support vector machine classifier and regressors.42 (7) Rempel, J. Y.; Bawendi, M. G.; Jensen, K. F. All SVM calculations were algorithmically tuned to pick the Insights into the Kinetics of Semiconductor best performing cost terms (all kernels) and gamma terms Nanocrystal Nucleation and Growth. J. Am. Chem. (radial basis function kernel) using “leave-one-out” cross Soc. 2009, 131, 4479–4489. validation as a performance metric. Nearest neighbors (8) Hamachi, L. S.; Jen-La Plante, I.; Coryell, A. C.; De regression was calculated using the FNN R package (version Roo, J.; Owen, J. S. Kinetic Control over CdS 1.1.2.1) with a k value of 3.43 Ridge regression was performed Nanocrystal Nucleation Using a Library of with the glmnet R package (version 2.0-16) for penalty Thiocarbonates, Thiocarbamates, and Thioureas. 44 constants (λ) of 0.001 to 50. Chem. Mater. 2017, 29, 8711–8719. ASSOCIATED CONTENT (9) Huang, H.; Raith, J.; Kershaw, S. V.; Kalytchuk, S.; Tomanec, O.; Jing, L.; Susha, A. S.; Zboril, R.; Supporting Information. The supporting Information is available free of charge on the ACS Publications Website. Rogach, A. L. Growth Mechanism of Strongly Photoluminescence emission wavelength as a function of Emitting CH3NH3PbBr3perovskite Nanocrystals particles size, photoluminescence emission fitting including with a Tunable Bandgap. Nat. Commun. 2017, 8 residuals, additional TEM images of nanoplatelets, modeling error (996), 1–7. as a function of bulk particle approximation, full tables of thickness (10) Andaraarachchi, H. P.; Thompson, M. J.; White, M. measurements gathered from integration of PL emission spectra, A.; Fan, H. J.; Vela, J. Phase-Programmed values used as alkylamine properties for t-test, full tables of Nanofabrication: Effect of Organophosphite wrapper results, and modeling error as a function of n-folds are Precursor Reactivity on the Evolution of Nickel and shown in Figures S1-S6 and Tables S1-S7 as well as video of 51 ACS Paragon Plus Environment 11


(11) (12)

(13)

(14)

(15)

(16)

(17) (18)

(19)

(20)

(21)

(22)

(23)

12

Page 12 of 15

Nickel Phosphide Nanocrystals. Chem. Mater. 2015, (24) Hachmann, J.; Olivares-amaya, R.; Atahan-evrenk, 27, 8021–8031. S.; Amador-bedolla, C.; Gold-parker, A.; Vogt, L.; Owen, J. The Coordination Chemistry of Brockway, A. M. The Harvard Clean Energy Nanocrystal Surfaces. Science. 2015, 347 (6222), Project : Large-Scale Computational. J. Phys. Chem. 615–616. Lett. 2011, 2, 2241–2251. Braun, M. B.; Korala, L.; Kephart, J. M.; Prieto, A. (25) Atif, M.; Afzal, F.; Cheng, C.; Hachmann, J. L. Synthetic Control of Quinary Nanocrystals of a Combining First-Principles and Data Modeling for Photovoltaic Material: The Clear Role of Chalcogen the Accurate Prediction of the Refractive Index of Ratio on Light Absorption and Charge Transport for Organic Polymers Combining First-Principles and Cu2– xZn1+ xSn(S1– ySey). ACS Appl. Energy Mater. Data Modeling for the Accurate Prediction of the 2018, 1, 1053-1059 Refractive Index of Organic Polymers. J. Chem. Lu, H.; Zhou, Z.; Prezhdo, O. V.; Brutchey, R. L. Phys. 2018, 148, 241712. Exposing the Dynamics and Energetics of the N(26) Dong, Y.; Qiao, T.; Kim, D.; Parobek, D.; Rossi, D.; Heterocyclic Carbene-Nanocrystal Interface. J. Am. Son, D. H. Precise Control of Quantum Chem. Soc. 2016, 138, 14844–14847. Confinement in Cesium Lead Halide Perovskite Depner, S. W.; Cultrara, N. D.; Farley, K. E.; Qin, Quantum Dots via Thermodynamic Equilibrium. Y.; Banerjee, S. Ferroelastic Domain Organization Nano Lett. 2018, 18, 3716–3722. and Precursor Control of Size in Solution-Grown (27) Almeida, G.; Goldoni, L.; Akkerman, Q.; Dang, Z.; Hafnium Dioxide Nanorods. ACS Nano 2014, 8, Khan, A. H.; Marras, S.; Moreels, I.; Manna, L. Role 4678–4688. of Acid-Base Equilibria in the Size, Shape, and Rondinone, A. J.; Samia, A. C. S.; Zhang, Z. J. A Phase Control of Cesium Lead Bromide Chemometric Approach for Predicting the Size of Nanocrystals. ACS Nano 2018, 12, 1704–1711. Magnetic Spinel Ferrite Nanoparticles from the (28) Bekenstein, Y.; Koscher, B. A.; Eaton, S. W.; Yang, Synthesis Conditions. J. Phys. Chem. B 2000, 104, P.; Alivisatos, A. P. Highly Luminescent Colloidal 7919–7922. Nanoplates of Perovskite Cesium Lead Halide and Their Oriented Assemblies. J. Am. Chem. Soc. 2015, Burrows, N. D.; Harvey, S.; Idesis, F. A.; Murphy, 137, 16008–16011. C. J. Understanding the Seed-Mediated Growth of (29) Tyagi, P.; Arveson, S. M.; Tisdale, W. A. Colloidal Gold Nanorods through a Fractional Factorial Organohalide Perovskite Nanoplatelets Exhibiting Design of Experiments. Langmuir 2017, 33, 1891– Quantum Confinement. J. Phys. Chem. Lett. 2015, 1907. 6, 1911–1916. Oliynyk, A. O.; Mar, A. Discovery of Intermetallic (30) Wang, H.; Whittaker-Brooks, L.; Fleming, G. R. Compounds from Traditional to Machine-Learning Exciton and Free Charge Dynamics of Approaches. Acc. Chem. Res. 2018, 51, 59–68. Methylammonium Lead Iodide Perovskites Are Oliynyk, A. O.; Adutwum, L. A.; Rudyk, B. W.; Different in the Tetragonal and Orthorhombic Pisavadia, H.; Lotfi, S.; Hlukhyy, V.; Harynuk, J. J.; Phases. J. Phys. Chem. C 2015, 119, 19590–19595. Mar, A.; Brgoch, J. Disentangling Structural (31) Di, D.; Musselman, K. P.; Li, G.; Sadhanala, A.; Confusion through Machine Learning: Structure Ievskaya, Y.; Song, Q.; Tan, Z. K.; Lai, M. L.; Prediction and Polymorphism of Equiatomic MacManus-Driscoll, J. L.; Greenham, N. C.; et al. Ternary Phases ABC. J. Am. Chem. Soc. 2017, 139, Size-Dependent Photon Emission from 17870–17881. Niemeyer, Z. L.; Milo, A.; Hickey, D. P.; Sigman, Organometal Halide Perovskite Nanocrystals M. S. Parameterization of Phosphine Ligands Embedded in an Organic Matrix. J. Phys. Chem. Reveals Mechanistic Pathways and Predicts Lett. 2015, 6, 446–450. Reaction Outcomes. Nat. Chem. 2016, 8, 610–617. (32) Dong, Q.; Fang, Y.; Shao, Y.; Mulligan, P.; Qiu, J.; De Jesus, L. R.; Stein, P.; Andrews, J. L.; Luo, Y.; Cao, L.; Huang, J. Electron-Hole Diffusion Lengths Xu, B. X.; Banerjee, S. Striping Modulations and > 175 Μm in Solution-Grown CH3NH3PbI3 Single Strain Gradients within Individual Particles of a Crystals. Science. 2015, 347 (6225), 967–970. Cathode Material upon Lithiation. Mater. Horizons (33) Shi, D.; Adinolfi, V.; Comin, R.; Yuan, M.; 2018, 5, 486–498. Alarousu, E.; Buin, A.; Chen, Y.; Hoogland, S.; Zhuo, Y.; Mansouri Tehrani, A.; Brgoch, J. Rothenberger, A.; Katsiev, K.; et al. Low Trap-State Predicting the Band Gaps of Inorganic Solids by Density and Long Carrier Diffusion in Organolead Machine Learning. J. Phys. Chem. Lett. 2018, 9, Trihalide Perovskite Single Crystals. Science. 2015, 1668–1673. 347 (6221), 519–522. Oliynyk, A. O.; Adutwum, L. A.; Harynuk, J. J.; (34) Swarnkar, A.; Chulliyil, R.; Ravi, V. K.; Irfanullah, Mar, A. Classifying Crystal Structures of Binary M.; Chowdhury, A.; Nag, A. Colloidal CsPbBr3 Compounds AB through Cluster Resolution Feature Perovskite Nanocrystals: Luminescence beyond Selection and Support Vector Machine Analysis. Traditional Quantum Dots. Angew. Chemie 2015, Chem. Mater. 2016, 28, 6672–6681. 127, 15644–15648. Pilania, G.; Mannodi-Kanakkithodi, A.; Uberuaga, (35) Protesescu, L.; Yakunin, S.; Bodnarchuk, M. I.; B. P.; Ramprasad, R.; Gubernatis, J. E.; Lookman, Krieg, F.; Caputo, R.; Hendon, C. H.; Yang, R. X.; T. Machine Learning Bandgaps of Double Walsh, A.; Kovalenko, M. V. Nanocrystals of Perovskites. Sci. Rep. 2016, 6 (19375), 1–10. Cesium Lead Halide Perovskites (CsPbX3, X = Cl, ACS Paragon Plus Environment

Page 13 of 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60

(36)

(37)

(38)

(39)

(40)

(41)

(42)

(43) (44)

(45)

(46) (47)

13

Chemistry of Materials Br, and I): Novel Optoelectronic Materials Showing (48) Cloutier, S. G. Impact of Different Surface Ligands Bright Emission with Wide Color Gamut. Nano Lett. on the Optical Properties of PbS Quantum Dot 2015, 15, 3692–3696. Solids. Materials. 2015, 8, 1858–1870. Huang, H.; Polavarapu, L.; Sichert, J. A.; Susha, A. (49) Choi, J. J.; Luria, J.; Hyun, B. R.; Bartnik, A. C.; S.; Urban, A. S.; Rogach, A. L. Colloidal Lead Sun, L.; Lim, Y. F.; Marohn, J. A.; Wise, F. W.; Halide Perovskite Nanocrystals: Synthesis, Optical Hanrath, T. Photogenerated Exciton Dissociation in Properties and Applications. NPG Asia Mater. 2016, Highly Coupled Lead Salt Nanocrystal Assemblies. 8, e328. Nano Lett. 2010, 10, 1805–1811. Akkerman, Q. A.; Motti, S. G.; Srimath Kandada, A. (50) Liang, Z.; Zhao, S.; Xu, Z.; Qiao, B.; Song, P.; Gao, R.; Mosconi, E.; D’Innocenzo, V.; Bertoni, G.; D.; Xu, X. Shape-Controlled Synthesis of AllMarras, S.; Kamino, B. A.; Miranda, L.; De Angelis, Inorganic CsPbBr3Perovskite Nanocrystals with F.; et al. Solution Synthesis Approach to Colloidal Bright Blue Emission. ACS Appl. Mater. Interfaces Cesium Lead Halide Perovskite Nanoplatelets with 2016, 8, 28824–28830. Monolayer-Level Thickness Control. J. Am. Chem. (51) Du, X.; Wu, G.; Cheng, J.; Dang, H.; Ma, K.; Zhang, Soc. 2016, 138, 1010–1016. Y. W.; Tan, P. F.; Chen, S. High-Quality CsPbBr3 Cho, J.; Jin, H.; Sellers, D.; Watson, D. F.; Son, D. perovskite Nanocrystals for Quantum Dot LightH.; Banerjee, S. Influence of Ligand Shell Ordering Emitting Diodes. RSC Adv. 2017, 7, 10391–10396. on Dimensional Confinement of Cesium Lead (52) Ulman, A. Formation and Structure of SelfBromide (CsPbBr3) Perovskite Nanoplatelets. J. Assembled Monolayers. Chem. Rev. 1996, 96, Mater. Chem. C 2017, 5, 8810–8818. 1533–1554. Lignos, I.; Stavrakis, S.; Nedelcu, G.; Protesescu, L.; (53) Silberzan, P.; Léger, L.; Ausserré, D.; Benattar, J. J. Demello, A. J.; Kovalenko, M. V. Synthesis of Silanation of Silica Surfaces. A New Method of Cesium Lead Halide Perovskite Nanocrystals in a Constructing Pure or Mixed Monolayers. Langmuir Droplet-Based Microfluidic Platform: Fast 1991, 7, 1647–1651. Parametric Space Mapping. Nano Lett. 2016, 16, (54) Cho, J.; Banerjee, S. Ligand-Directed Stabilization 1869–1877. of Ternary Phases: Synthetic Control of Structural Shamsi, J.; Dang, Z.; Bianchini, P.; Canale, C.; Di Dimensionality in Solution-Grown Cesium Lead Stasio, F.; Brescia, R.; Prato, M.; Manna, L. Bromide Nanocrystals. Chem. Mater. 2018, 30, Colloidal Synthesis of Quantum Confined Single 6144–6155. Crystal CsPbBr3 Nanosheets with Lateral Size (55) Mistry, A.; Juarez-Robles, D.; Stein, M.; Smith, K.; Control up to the Micrometer Range. J. Am. Chem. Mukherjee, P. P. Analysis of Long-Range Soc. 2016, 138, 7240–7243. Interaction in Lithium-Ion Battery Electrodes. J. Liu, Z.; Bekenstein, Y.; Ye, X.; Nguyen, S. C.; Electrochem. Energy Convers. Storage 2016, 13, Swabeck, J.; Zhang, D.; Lee, S. T.; Yang, P.; Ma, 031006. W.; Alivisatos, A. P. Ligand Mediated (56) Viswanatha, R.; Santra, P. K.; Dasgupta, C.; Sarma, Transformation of Cesium Lead Bromide Perovskite D. D. Growth Mechanism of Nanocrystals in Nanocrystals to Lead Depleted Cs4PbBr6 Solution: ZnO, a Case Study. Phys. Rev. Lett. 2007, Nanocrystals. J. Am. Chem. Soc. 2017, 139, 5309– 98, 98–101. 5312. (57) Johnson, N. J. J.; Korinek, A.; Dong, C.; Van Meyer, D.; Dimitriadou, E.; Hornik, K.; Weingessel, Veggel, F. C. J. M. Self-Focusing by Ostwald A.; Leisch, F. E1071: Misc Functions of the Ripening: A Strategy for Layer-by-Layer Epitaxial Department of Statistics, Probability Theory Group Growth on Upconverting Nanocrystals. J. Am. (Formerly: E1071), TU Wien. 2017. Chem. Soc. 2012, 134, 11068–11071. Alina, B.; Sham, K.; John, L.; Sunil, A.; David, M.; (58) Hoffman, J. B.; Zaiats, G.; Wappes, I.; Kamat, P. V. Shengqiao, L. FNN: Fast Nearest Neighbor Search CsPbBr3Solar Cells: Controlled Film Growth Algorithms and Applications. CRAN 2018. through Layer-by-Layer Quantum Dot Deposition. Friedman, J.; Hastie, T.; Tibshirani, R. Chem. Mater. 2017, 29, 9767–9774. Regularization Paths for Generalized Linear Models (59) ACD/Percepta. Advanced Chemistry Development via Coordinate Descent. J. Stat. Softw. 2010, 33, 1Inc.: Toronto, ON 2017. 22. (60) Haynes, William M., David R. Lide, and T. J. B. CRC Handbook of Chemistry and Physics : A Cho, J.; Choi, Y. H.; O’Loughlin, T. E.; De Jesus, Ready-Reference Book of Chemical and Physical L.; Banerjee, S. Ligand-Mediated Modulation of Data, 97th ed.; Boca Raton, Florida, 2017. Layer Thicknesses of Perovskite Methylammonium Lead Bromide Nanoplatelets. Chem. Mater. 2016, (61) National Institute of Advanced Industrial Science 28, 6909–6916. and Technology, Spectral Database for Organic Cottingham, P.; Brutchey, R. L. On the Crystal Compounds, http://sdbs.db.aist.go.jp, acsessed July Structure of Colloidally Prepared CsPbBr3 Quantum 25, 2017 (62) Matulis, D.; Bloomfield, V. A. Thermodynamics of Dots. Chem. Commun. 2016, 52, 5246–5249. the Hydrophobic Effect. II. Calorimetric Cottingham, P.; Brutchey, R. L. Compositionally Measurement of Enthalpy, Entropy, and Heat Dependent Phase Identity of Colloidal CsPbBr3Capacity of Aggregation of Alkylamines and Long I Quantum Dots. Chem. Mater. 2016, 28 (21), X x Aliphatic Chains. Biophys. Chem. 2001, 93, 53–65. 7574–7577. ACS Paragon Plus Environment


(63)

Hanwell, M. D.; Curtis, D. E.; Lonie, D. C.; Vandermeerschd, T.; Zurek, E.; Hutchison, G. R. Avogadro: An Advanced Semantic Chemical Editor, Visualization, and Analysis Platform. J.

Cheminform. 2012, 4 (17), 1-17.


Page 14 of 15

Page 15 of 15 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60


ToC Graphic


15

Machine Learning-Directed Navigation of Synthetic Design Space: A

Recommend Documents