Machine-learning-optimized perovskite nanoplatelet synthesis Carola Lampe1Ioannis Kouroudis2Milan Harth2

2025-05-02 0 0 869.3KB 14 页 10玖币

侵权投诉

Machine-learning-optimized perovskite

nanoplatelet synthesis

Carola Lampe1Ioannis Kouroudis2Milan Harth2

Stefan Martin1Alessio Gagliardi2Alexander S. Urban1

1Nanospectroscopy Group and Center for NanoScience,

Nano-Institute Munich, Faculty of Physics

2Department of Electrical and Computer Engineering, Technical

University of Munich

Abstract

With the demand for renewable energy and eﬃcient devices rapidly

increasing, a need arises to ﬁnd and optimize novel (nano)materials. This

can be an extremely tedious process, often relying signiﬁcantly on trial

and error. Machine learning has emerged recently as a powerful alter-

native; however, most approaches require a substantial amount of data

points, i.e., syntheses. Here, we merge three machine-learning models with

Bayesian Optimization and are able to dramatically improve the quality

of CsPbBr3nanoplatelets (NPLs) using only approximately 200 total syn-

theses. The algorithm can predict the resulting PL emission maxima of

the NPL dispersions based on the precursor ratios, which lead to previ-

ously unobtainable 7 and 8 ML NPLs. Aided by heuristic knowledge, the

algorithm should be easily applicable to other nanocrystal syntheses and

signiﬁcantly help to identify interesting compositions and rapidly improve

their quality.

1 Introduction

Halide perovskite nanocrystals (PNCs), ﬁrst demonstrated in 2014, have been

rapidly improved, yielding tunability throughout the visible spectrum, quantum

arXiv:2210.09783v1 [physics.app-ph] 18 Oct 2022

Figure 1: Scheme of the optimization process: Initially existing data points (syn-

theses) were analyzed and used to predict a spectral ﬁgure of merit (FoM) based

on the narrowness and symmetry of their PL spectra using Gaussian Processes

in combination with a Random Forest and a Neural Network. Constrained

Bayesian optimization subsequently leverages the information of the artiﬁcial

intelligence step to provide suggestions for new compositions. For each cycle,

14 new syntheses were carried out and characterized. The amended dataset is

subjected to a new optimization cycle.

yields approaching 100%, and diverse geometries and sizes.Schmidt et al. [2014]

Due to their exceptional properties, PNCs have already been incorporated into

diverse applications, focusing on optoelectronics such as LEDs, solar cells, and

photodetectors, but also in ﬁeld-eﬀect transistors and, even more recently, pho-

tocatalysis. Shamsi et al. [2019a], Abiram et al. [2022], Zhang et al. [2017],

Gao et al. [2017], Shyamal et al. [2020] Despite these impressive improvements,

several issues impede widespread commercialization, such as stability, lead toxi-

city, and spectral eﬃciency in the blue region of the visible spectrum.Schoonman

[2015], Schileo and Grancini [2021], Weng et al. [2022] This latter eﬀect is due

to the chloride-perovskites being far from defect tolerant, resulting in extremely

poor eﬃciencies compared to bromide- and iodide-based perovskites. Ye et al.

[2021] Another way to tune the spectral response in PNCs is through quantum

conﬁnement. Especially, 2D nanoplatelets (NPLs) are ideal in this regard, as

they exhibit no inhomogeneous broadening in the conﬁned dimension, with only

incremental thickness values possible - currently between 2 and 6 monolayers

(MLs). Bohn et al. [2018] Analogous to the bulk-like Ruddlesden-Popper per-

ovskitesStoumpos et al. [2016], their strong conﬁnement can enable directional

emission, boosting maximum external quantum eﬃciencies to 28%.Morgenstern

et al. [2020] The quality of these colloidal quantum wells has improved signiﬁ-

cantly; however, the quantum yields are still far from unity, and reproducibility

is an issue. Improving the NPL quality or that of any NC dispersion is an ardu-

ous task, involving a vast possible number of parameters relating to composition

and fabrication. Synthesizing all of these is both infeasible and unnecessary, as it

is possible to create robust and data-eﬃcient predictors to describe the outcome

of changes in fabrication parameters.Mayr et al. [2022] Artiﬁcial neural networks

(ANNs) have been widely used to approximate the quantities of interest,Chen

and Pao [2019], Regonia et al. [2020] but success has also been achieved with,

e.g., random forests and support vector machines. Baum et al. [2020], Rickert

et al. [2021] It seems clear that no algorithm is universally superior, but rather

each may be more suited for speciﬁc applications.Wolpert and Macready [1997]

Despite their impressive results, these methods occasionally suﬀer from extrap-

olation into areas of input space with sparse or no data. Therefore, material

science is a fertile ground for applying methods with inbuilt uncertainty quan-

tiﬁcation functionalities, such as Gaussian processes (GPs), which are especially

attractive given their data eﬃciency and robustness against overﬁtting.Li et al.

[2020], Zhang and Xu [2020], Seko and Ishiwata [2020] These methods provide

an excellent set of predictors to determine the eﬀect that diﬀerent fabrication

and chemical parameters have on the resulting material properties. Neverthe-

less, the optimal values of these parameters are not trivially determined. This

ﬁeld has lately been dominated by Bayesian optimization. In this scheme, the

value of the predicted objective function is weighted against the intrinsic un-

certainty of the prediction to balance the exploration of new areas against the

exploitation of already acquired information.Balachandran et al. [2018], Voznyy

et al. [2019]. However, this has required a massive experimental eﬀort to achieve

impressive results. In the future, models to optimize novel materials must fo-

cus on two aspects: incorporating multiple properties to increase the versatility

of the targeted materials and reducing the number of syntheses necessary to

achieve good results.

In this study, we develop a model addressing both of these issues as illus-

trated in Figure 1. By combining Gaussian processing with a neural network

and a random forest classiﬁer, we can signiﬁcantly reduce the demand for op-

timizing the synthesis of 2D CsPbBr3-based NPLs. The algorithm uses three

precursor amounts to predict the emission wavelength of the resulting NPLs

and the quality, i.e., the homogeneity of the photoluminescence (PL) spectra.

Starting from a pool of 100 initial syntheses, we carried out seven rounds of op-

timization. The algorithm produced 14 new precursor combinations, which were

then used for synthesis and the PL of the resulting dispersions measured. For

all previously synthesized NPL thicknesses (2-6 ML), we signiﬁcantly reduced

the width and asymmetry of the PL emission, signifying higher homogeneity.

Additionally, the algorithm eﬀectively predicted precursor combinations leading

to hitherto unobtained, thicker NPLs (7 and 8 MLs). The algorithm’s perfor-

mance was exceptional, especially considering the small amount of necessary

experimental synthesizing.

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Machine-learning-optimizedperovskitenanoplateletsynthesisCarolaLampe1IoannisKouroudis2MilanHarth2StefanMartin1AlessioGagliardi2AlexanderS.Urban11NanospectroscopyGroupandCenterforNanoScience,Nano-InstituteMunich,FacultyofPhysics2DepartmentofElectricalandComputerEngineering,TechnicalUniversityofMunich...

展开>> 收起<<

Machine-learning-optimized perovskite nanoplatelet synthesis Carola Lampe1Ioannis Kouroudis2Milan Harth2.pdf

共14页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Machine-learning-optimized perovskite nanoplatelet synthesis Carola Lampe1Ioannis Kouroudis2Milan Harth2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: