In Situ Photoluminescence Imaging Dataset of Blade-Coated Perovskite Photovoltaics

<strong>Content:</strong> The dataset contains time-resolved in situ images acquired during the formation of the perovskite layer which is then built into a perovskite solar cell. The image time series in the dataset encompass the drying and crystallization of the blade-coated perovskite thin-films. An initial exploration of the data presented in the dataset is conducted in the paper <strong>Process Insights into Perovskite Thin-Film Photovoltaics from Machine Learning with In Situ Luminescence Data.</strong> A total of 1,129 solar cells were fabricated using the blade coating deposition method. To monitor the vacuum quenching process of the perovskite layer, a photoluminescence (PL) imaging setup was used to capture four channels of image data. These channels included time series images (2D+t) captured through various spectral filters, with one channel showing reflectance and the other three showing different parts of the PL spectrum. The three PL channels with different spectral transmissions were also used to compute a image time series of spatially resolved PL peak wavelengths. All images were cropped into smaller patches of 65x56 pixels each, which only included the active area of a single solar cell. Different metrics are available as target variables. For each solar cell in the dataset, the photovoltaic performance parameters, namely (1) power conversion efficiency (PCE), (2) open-circuit voltage (<em>V_OC</em>), (3) short-circuit current density (<em>J_SC</em>), and (4) fill factor (FF)), are available (measured backward and forward, as well as the average between forward and backward). Furthermore, information about the perovskite layer thickness of each solar cell’s active area is provided: mean thickness, root-mean-square thickness, and peak-2-valley thickness. Also, additional information like substrate ID and the position of each solar cell within its substrate is provided. All solar cells were fabricated using the same materials, methods, and experimental parameters. As a result, the dataset can be used to apply machine learning techniques to identify variations in the fabrication process between iterations, improve understanding of the process, and predict performance in-line before completing the half-stack into a functional solar cell. Further information on the experimental acquisition procedure can be found in the paper <strong>Process Insights into Perovskite Thin-Film Photovoltaics from Machine Learning with In Situ Luminescence Data</strong>. <strong>Usage:</strong> The dataset is made available as a single hdf5-file. The npy-data can be extracted using the notebook “00_extract_data_from_hdf5_file.ipynb” which is provided in the GitHub repository https://github.com/AI-InSu-Pero/ML-PerovskitePV-InSituLuminescene The structure of the dataset after extraction from the hdf5-file is depicted below. The dataset (1,129 solar cells) is split into two subfolders, containing train (780 solar cells) and test data (349 solar cell), respectively. For training and test data, the corresponding labels are listed in csv files. In the train and test folders, there are subfolders for each of the substrate assigned to either of the two sets. In the substrate folders, the data for all the patches of a substrate is saved in npy-format with the shape (719, 5, 65, 56), representing (time step, channel, image height, image width). It can be loaded using numpy.load(path_to_file). The order of the five channels is as follows: (0) reflectance, (1) entire PL spectrum, (2) filtered PL spectrum – longer wavelengths remaining, (3) filtered PL spectrum – shorter wavelengths remaining, (4) computed peak wavelength of PL spectrum. In the train folder, an additional folder “cv_splits_5fold” gives the train and validation splits for the 5-fold cross-validation used in the dataset exploration paper. For each fold, the labels are given as csv-files for train and validation split. <pre><code>dataset ├── train │ ├── ACA │ │ ├── 11.npy │ │ ├── 12.npy │ │ ├── 13.npy │ │ ├── 14.npy │ │ ├── 21.npy │ │ └── ... (all other patches of this substrate) │ ├── ACA │ │ ├── 11.npy │ │ ├── 12.npy │ │ ├── 13.npy │ │ ├── 14.npy │ │ ├── 21.npy │ │ └── ... (all other patches of this substrate) │ ├── ... (all other train substrates) │ ├── cv_splits_5fold │ │ ├── fold0 │ │ │ ├── train.csv │ │ │ └── val.csv │ │ └── ... (all other folds) │ └─── labels.csv └── test ├── ACE │ ├── 11.npy │ ├── 12.npy │ ├── 13.npy │ ├── 14.npy │ ├── 21.npy │ └── ... (all other patches of this substrate) ├── ... (all other test substrates) └── labels.csv </code></pre>**内容：** 本数据集包含钙钛矿层成型过程中的时间分辨原位图像，该钙钛矿层最终将被制备为钙钛矿太阳能电池。数据集内的图像序列涵盖刮刀涂布制备的钙钛矿薄膜的干燥与结晶过程。本数据集相关数据的初步探索研究已发表于论文**《基于原位发光数据结合机器学习的钙钛矿薄膜光伏器件工艺解析》(Process Insights into Perovskite Thin-Film Photovoltaics from Machine Learning with In Situ Luminescence Data)**。 本研究共采用刮刀涂布沉积法制备了1129个太阳能电池。为监测钙钛矿层的真空淬火工艺，研究人员采用光致发光（photoluminescence, PL）成像装置采集了四通道图像数据。该四通道数据包含通过不同光谱滤光片采集的时间序列图像（2D+t）：其中一个通道为反射率图像，其余三个通道分别采集光致发光光谱的不同波段信息。三个具有不同光谱透射特性的PL通道数据还可用于计算空间分辨的PL峰波长时间序列图像。所有图像均被裁剪为65×56像素的图像块，仅包含单个太阳能电池的活性区域。 数据集提供多种指标作为目标变量。针对每个太阳能电池，其光伏性能参数包括：(1) 功率转换效率（power conversion efficiency, PCE）、(2) 开路电压（open-circuit voltage, V_OC）、(3) 短路电流密度（short-circuit current density, J_SC）以及(4) 填充因子（fill factor, FF），上述参数均提供正反扫描测量值以及正反扫平均值。此外，数据集还提供每个太阳能电池活性区的钙钛矿层厚度信息：平均厚度、均方根厚度以及峰谷厚度。同时还包含衬底ID以及每个电池在衬底上的位置信息。 所有太阳能电池均采用相同的材料、制备方法与实验参数制备。因此，该数据集可用于开展机器学习相关研究，以识别不同批次制备工艺间的差异，加深对制备工艺的理解，并可在完成半叠层制备为功能太阳能电池前实现在线性能预测。关于实验采集流程的更多细节可参阅上述论文。 **使用说明：** 本数据集以单个HDF5文件形式提供。可通过GitHub仓库https://github.com/AI-InSu-Pero/ML-PerovskitePV-InSituLuminescene 中提供的Jupyter Notebook“00_extract_data_from_hdf5_file.ipynb”提取其中的Numpy（npy）格式数据。从HDF5文件提取后的数据集结构如下所示。 本数据集共包含1129个太阳能电池，分为训练集（780个）与测试集（349个）两个子文件夹。训练集与测试集的对应标签均存储于CSV文件中。在训练集与测试集文件夹下，分别设有对应各分配至对应数据集的衬底的子文件夹。在衬底文件夹中，单个衬底的所有图像块数据均以npy格式存储，数据形状为(719, 5, 65, 56)，分别对应（时间步数、通道数、图像高度、图像宽度），可通过numpy.load(path_to_file)加载。五个通道的顺序如下：(0) 反射率、(1) 全PL光谱、(2) 滤波后保留长波长的PL光谱、(3) 滤波后保留短波长的PL光谱、(4) 计算得到的PL光谱峰波长。 训练集文件夹下还设有额外的“cv_splits_5fold”文件夹，其中包含本数据集探索论文中使用的5折交叉验证的训练集与验证集划分方式。每个折次的训练集与验证集标签均以CSV文件形式提供。 数据集 ├── 训练集（train） │ ├── ACA │ │ ├── 11.npy │ │ ├── 12.npy │ │ ├── 13.npy │ │ ├── 14.npy │ │ ├── 21.npy │ │ └── ... （该衬底的其余所有图像块） │ ├── ACA │ │ ├── 11.npy │ │ ├── 12.npy │ │ ├── 13.npy │ │ ├── 14.npy │ │ ├── 21.npy │ │ └── ... （该衬底的其余所有图像块） │ ├── ... （所有其余训练集衬底） │ ├── cv_splits_5fold │ │ ├── fold0 │ │ │ ├── train.csv │ │ │ └── val.csv │ │ └── ... （其余所有折次） │ └── labels.csv └── 测试集（test） ├── ACE │ ├── 11.npy │ ├── 12.npy │ ├── 13.npy │ ├── 14.npy │ ├── 21.npy │ └── ... （该衬底的其余所有图像块） ├── ... （所有其余测试集衬底） └── labels.csv