The local chemical composition of halide perovskites is a crucial factor in determining their macroscopic properties and their stability. While the combination of scanning transmission electron microscopy (STEM) and energy-dispersive X-ray spectroscopy (EDX) is a powerful and widely used tool for accessing such information, electron-beam-induced damage and complex formulation of the films make this investigation challenging. Here we demonstrate how multivariate analysis -including statistical

more »

... utines derived from "big data" research, such as Principal Component Analysis, PCA -can be used to dramatically improve the signal recovery from fragile materials. We also show how a similar decomposition algorithm (Non-negative Matrix Factorisation, NMF) can unravel elemental composition at the nanoscale in perovskite films, highlighting the presence of segregated species and identifying the local stoichiometry at the nanoscale. Research into hybrid organic-inorganic perovskite solar cells (PSCs) has flourished over recent years, attracting strong interest by the scientific community 1-3 . This emerging class of devices has become increasingly popular due to the opportunity of reaching good power conversion while being compatible with wet chemistry processing for large area devices. The global research effort associated with the rise of hybrid perovskites resulted in rapid advances in chemical formulation 4-7 , fabrication methods 8-10 and device architecture 11,12 -however, most of the progress has been through empirical device improvements, and a number of key questions still remain unanswered. Open issues are covered in reviews 13,14 and include the optimal chemical composition of the perovskite films 11 , ion migration 15,16 , scalable fabrication routes 17,18 , device architecture 19 and stability in operation 20 . Specifically, long term stability of the modules under operating conditions is considered to be the main drawback preventing commercial applications 21 . However, poor stability not only prevents easy commercialisation -it also complicates scientific research: commonly used optical and analytical characterisation tools can induce reversible or irreversible structural/chemical changes in the perovskite films through the use of high energy photon or electron beams [22] [23] [24] [25] . In this work we propose an approach that combines the acquisition of high resolution chemical maps by scanning transmission electron microscopy (STEM) with dedicated MultiVariate Analysis (MVA) 26 methods that improve signal/noise ratio (SNR) and identify correlations between the spatial distribution of elements. Such correlations emerge from statistical analysis, and consist of maps that describe the distribution of chemical compounds rather than just elements. The use of methods that minimise operator input improves the reproducibility of results and the sensitivity to unexpected chemical compounds, such as phases with unpredicted stoichiometry, or elemental segregation. This approach is particularly valuable for hybrid perovskite-based films and devices, in which complex compounds can form, ionic species are prone to migration, and the electron dose during STEM analysis needs to be minimised to prevent local damage. Specifically, we tested and compared different computational methods -Principal Component Analysis (PCA) 27 and Non-Negative Matrix Factorisation (NMF) 28 , demonstrating how they can be used to increase the SNR ratio of a dataset and provide new insights on the local chemistry. The process flow of the paper is as follows: Initially, a cross-sectional sample is extracted from a solar cell (or a perovskite film) using conventional Focused Ion Beam (FIB) preparation 29 and transferred to a (scanning) transmission electron microscope (STEM). STEM-EDX (Energy Dispersive X-ray spectroscopy) analysis is carried out using optimised illumination conditions to limit the electron dose on the sample, and the experimental data is processed using MVA algorithms available in open source scientific analysis packages such as python-based Hyperspy 30 . SEM, TEM and STEM are widely used techniques for nanomaterials characterisation. The electron beam, however, can cause temporary or permanent changes in the sample. Such changes can be related to local charging, heating, chemical reactions, preferential removal of atomic species or deposition of carbonaceous contamination 31 . These phenomena are dependent on several factors, including the atomic number of the elements in the specimen, their bonding, and the electrical and thermal conductivity of the specimen, as well as the primary beam illumination