Despite the remarkable optoelectronic efficiencies of hybrid perovskite based devices heterogeneity in both the performance and stability is observed at the nanoscale [1,2]. Furthermore, improvements in the performance of solar cells, LEDs and X-ray detectors are often achieved through empirical findings, with a full understanding of the underlying structural mechanisms yet to be elucidated. Considering this fact, we present how correlative multimodal microscopy, enabled by computer vision algorithms, has allowed for structural insights into: the intrinsic quantum confinement phenomena observed in FAPbI₃ [3]; and the origins of photo-induced halide segregation observed in mixed halide compositions [4].

Scanning electron diffraction (SED), a variant of 4D-STEM, allows for information on crystallographic phase, orientation, and defects to be uncovered; whereas hyperspectral photoluminescence (PL) mapping provides information on optoelectronic characteristics. The spatial correlation between the two techniques therefore provides a direct link between structure and performance. Similarly to the work of Jones, Osherov and Alsari et al. [5] to find common areas between separate tecniques, gold fiducial markers are synthesized and deposited onto a thin film. Due to the resolution differences between SED (typical probe size ∼5 nm) and hyperspectral PL mapping (resolution diffraction limited to ∼100s of nm) multiple contiguous SED scans are recorded before being stitched together during post processing. To do this robustly a keypoint detection and matching algorithm is applied to virtual bright field images created from the SED data [6]. Once the keypoints are detected the random sample consensus algorithm, or variants thereof, is used to define an affine transform which stitches one image onto the other, thus correcting for differences in rotation, shear, and translation [6]. To finally coregister the hyperspectral PL and stitched SED datasets several options are available, however common approaches operate on the principle of maximising metrics such as the normalised cross correlation or mutual information between datasets [7]. Importantly, as all image transforms are known we can then calculate where each ‘pixel’ from a SED scan corresponds to in the hyperspectral PL data.

Due to the size and complexity of the resulting datasets interpretation of the data is challenging. For instance, if only 50 SED scans are considered, each of which contains 512 x 512 diffraction patterns (typical during one experimental session), this amounts to a total of 13,107,200 individual patterns and hundreds of gigabytes of data. To mitigate this problem and rapidly reduce the dimensionality of the data we have adapted the simple linear iterative clustering (SLIC) algorithm, prevalent in the field of remote sensing [8]. This approach allows us to reduce hundreds of thousands of individual diffraction patterns to ∼100 single crystal patterns obtained by averaging over individual grains of the polycrystalline film. Remarkably, once parallelised, this approach proves exceptionally computationally efficient with typical compute times taking approximately a minute using a standard desktop machine (32Gb RAM, 11th Gen Intel(R) Core(TM) i5-11400 CPU). An automated indexing procedure of the clustered SED patterns can then be employed to obtain phase and orientation maps over large areas (typically ∼20x20 μm) at low computational cost.

This approach has allowed the structural causes of the intrinsic quantum confinement phenomena observed at cryogenic temperatures in FAPbI₃ to be attributed to {111}_c type nanoscale twinning, and an understanding of the structural origins of photo-induced halide segregation to be obtainable. We hope this toolkit, with data analysis pipelines which are open source, enables correlative microscopy experiments with other instruments and on other material systems. Furthermore, we anticipate the insights provided on hybrid perovskites will allow for more directed modifications to the fabrication of devices, thereby accelerating improvements in device efficiency and stability.