Solar Energy Materials & Solar Cells xxx (xxxx) xxx Please cite this article as: Çağla Odabaşı, Ramazan Yıldırım, Solar Energy Materials & Solar Cells, https://doi.org/10.1016/j.solmat.2019.110284 0927-0248/© 2019 Elsevier B.V. All rights reserved. Machine learning analysis on stability of perovskite solar cells Çagla Odabas¸ı, Ramazan Yıldırım * Department of Chemical Engineering, Bogaziçi University, 34342, Bebek, Istanbul, Turkey A R T I C L E INFO Keywords: Perovskite solar cells Data mining Machine learning Association rule mining Stability Knowledge extraction ABSTRACT In this work, a dataset containing long-term stability data for 404 organolead halide perovskite cells was con- structed from 181 published papers and analyzed using machine-learning tools of association rule mining and decision trees; the effects of cell manufacturing materials, deposition methods and storage conditions on cell stability were investigated. For regular cells, mixed cation perovskites, multi-spin coating as one-step deposition, DMF DMSO as precursor solution and chlorobenzene as anti-solvent were found to have positive effects on stability; SnO 2 as ETL compact layer, PCBM as second ETL, inorganic HTLs or HTL-free cells, LiTFSI TBP FK209 and F4TCNQ as HTL additives and carbon as back contact were also found to improve stability. The cells stored under low humidity were found to be more stable as expected. The degradation was slightly faster in inverted cells under humid condition; the use of some materials (like mixed cation perovskites, PTAA and NiO x as HTL, PCBM C60 as ETL, and BCP interlayer) were found to result in more stable cells. 1. Introduction Organolead halide perovskite solar cells (PSCs) have been attracted great attention in recent years, and the power conversion efficiency (PCE) has reached to 23.7% in less than 10 years [1]. However, the challenges in long term stability remained unsolved preventing the commercialization of this promising technology because the efficiency, stability and cost (golden triangle) are the basic requirement for any practical application [2]. Consequently, there has been a considerable shift in research focus to stability; indeed, the number of research arti- cles involving the stability has been increased in recent years as evident from Fig. 1 (from Web of Science search with the keyword search of perovskite solar and stability in topic segment on April 03, 2019). There has been also considerable progress in the stability of PSCs; for example, Arora et al. [3] and Christians et al. [4] reported 1000 h stability under illumination while the cells retaining 95% of the initial cell efficiency by modifying their HTL. Similarly, Grancini et al. [5] achieved stability of solar modules more than 10000 h without PCE loss using 2D/3D perovskites. There are large numbers of publications covering the long-term stability of the perovskite cells. These publications usually report the effects of individual materials (like individual perovskite) and deposi- tion methods (like one or two step) as well as the storage conditions (light, humidity, temperature and oxygen) used in those works; there are also reports that compare the effects of material alternatives investigated in the same work or in various publications. Consequently, there is a massive accumulation of data, which likely contain invaluable information on this subject; however, it is not easy to utilize this collection of data by naked eyes because it is too complex (too many variables with too many options), non-uniform and scattered over a large number of publications. Sufficiently large datasets can be con- structed from the related publications and can be analyzed using ma- chine learning tools to identify the general patterns, trends and significant factors; then, the results obtained to understand the experi- ence in the field can help improving stability further. Indeed, we have implemented this approach in various fields of energy such as water gas shift reaction [6], CO oxidation [7], CO 2 adsorption [8], dry reforming of methane [9], biodiesel production [10], and water splitting [11], and obtained significant success in making some valuable generalization. We recently published a work involving the efficiency analysis of perovskite solar cells using machine learning tools [12]; there are also some other recent attempts to understand the various aspect of this potentially important technology using machine learning [13–17]. However, as far as we know, there is no published work to understand the effects of cell manufacturing materials and methods on cell stability by analyzing the results of large number of published experimental works performed for this purpose. In this work, we constructed a dataset containing the stability profiles (power conversion efficiency versus time data) of 404 cells (from 181 publications), which were manufac- tured and tested under various conditions. Then, we analyzed the * Corresponding author. E-mail address: yildirra@boun.edu.tr (R. Yıldırım). Contents lists available at ScienceDirect Solar Energy Materials and Solar Cells journal homepage: http://www.elsevier.com/locate/solmat https://doi.org/10.1016/j.solmat.2019.110284 Received 1 August 2019; Received in revised form 18 October 2019; Accepted 7 November 2019