The paper presents a way to let a microscope and machine learning work together to find how tiny structural features control material behavior. The authors built an autonomous workflow that steers a microscope to new and interesting places and then learns links between local structure and local electrical response. They tested the approach on hybrid (halide) perovskite thin films and found clear connections between nanoscale geometry and electrical hysteresis — where a device’s current depends on the history of applied voltage.

To do this the team combined two machine-learning tools. First, a dual-novelty deep kernel learning (DN-DKL) method decides where the microscope should take the next measurement. DN-DKL scores novelty both for local spectra and for structural images so the system prefers places that look different from what it has already seen. Second, a dual variational autoencoder (VAE) learns compact representations of images and spectra and embeds them together in a shared “map” or latent space. The instrument used for experiments was conductive atomic force microscopy (C-AFM), which gives both topography and local current–voltage curves.

At a high level, novelty-driven acquisition gathers a diverse and large set of local spectroscopic measurements. The shared latent map from the dual-VAE then groups locations that have similar structure and similar spectroscopic responses. This multimodal representation acts as a structure–property relationship map. Human experts then inspect that map to extract physical insight and to connect machine-found patterns to known material physics.

The authors report that this combined approach reveals distinct electrical regimes tied to specific nanoscale motifs. They resolved differences among grain interiors, grain-boundary grooves, and triple-junction nano-traps. They found that hysteresis and charge transport are controlled more by localized geometric trap structures than by average grain size. In particular, grain-boundary junction points showed hysteresis under different bias conditions, and asymmetric grain boundaries were found to suppress charge transport.