Scanning Electron Diffraction and Unsupervised Methods: a Study on Photovoltaic Absorbers

Several candidate materials and structures show great promise for next generation photovoltaics and optoelectronic devices. The complete understanding of charge dynamics at the nanoscale, and in particular how local composition and crystallography affect transport phenomena, recombination and ultimately power conversion efficiency, is often a bottleneck in the development of architectures that can be deployed and have an impact outside of the research lab.

Transmission electron microscopy enables studies of elemental distributions as well as local morphology, with cross-sectional observation revealing variations through the depth of the absorber layer as well as quality of interfaces across all components of a device stack. Furthermore, recent developments in technique, hardware and software now enable the acquisition of electron diffraction patterns across micron-scale areas with a probe size of a few nm, yielding very local information on crystal phase and orientation. This approach, often referred to as 4DSTEM, generates very large, information-rich datasets. Their interpretation can be carried out manually, with careful analysis of selected regions, or through the use of unsupervised routines that extract information via the application of statistical methods.

In this study we looked at three candidate absorbers for next generation photovoltaics: evaporated halide perovskites, Cu(InGa)S2 (CIGS) and Sb2Se3. While they all rely on optimisation of fabrication parameters in order to obtain the best optoelectronic parameters, each case presents different challenges. Halide perovskites are rapidly damaged by the electron beam, and need to be studied with a very low electron dose. CIGS combine a complex stoichiometry, requiring a fine tuning of the compositional gradient through the thickness, as well as featuring different types of grain boundaries, some of which are particularly detrimental to electron transport. Finally, Sb2Se3 films comprise a large number of small (tens of nm) crystallites, requiring fine sampling and computationally efficient routines to analyse the corresponding 4DSTEM datasets.

We employ a variety of unsupervised methods for extracting information from the datasets acquired on these systems: Non-negative Matrix Factorization, K-means clustering and Mini-batch K-means clustering. The strengths of each approach are highlighted and discussed.

The application of 4DSTEM to materials for optoelectronics is a very versatile tool that can address issues in a variety of promising candidate materials. With this work we demonstrate a framework for compositional and crystallographic studies at single-grain scale, with potential applications in new, emerging semiconductors.