Closing the Feedback Loop: Live Polarization Analysis for Automated In Situ 4D-STEM and SEM

Ferroelectric devices are dynamically evolving systems whose functionality emerges from coupled polarization processes spanning atomic to micron length scales [1]. In operating devices, polarization continuously reconfigures through domain nucleation, growth, wall motion, and interfacial charge redistribution. Subtle picometer scale atomic displacements at interfaces and domain walls govern mesoscale domain architecture, which in turn dictates macroscopic switching behavior, conductivity pathways, and overall device performance. Domains are elastically, electrostatically, and electronically interconnected and do not act independently. Consequently, small atomic scale shifts can propagate across scales, altering polarization dynamics and functional response in real time. Understanding ferroelectric devices therefore requires a unified multiscale framework that captures structural hierarchy and temporal evolution under electrical bias. Modern electron microscopy provides access to this hierarchy through complementary techniques including scanning electron microscopy, scanning transmission electron microscopy, and four dimensional STEM. While established polarization mapping methods in STEM rely on analytical tracking of atomic column displacements, comparable generalizable approaches for SEM and 4D STEM that provide rapid feedback remain limited. Offline 4D STEM analysis workflows typically require dataset specific configuration and extensive post processing, which constrains their use during in situ experiments. In the context of dynamic biasing studies, where large multidimensional datasets are acquired rapidly, real time interpretation is essential for informed experimental decision making. To address this challenge, we build upon our previously established automated in situ biasing workflows for 4D STEM and EELS acquisition [2] and integrate machine learning directly at the microscope to enable live polarization analysis. Automated biasing sequences coordinate electrical stimulation with synchronized diffraction and spectroscopy acquisition, ensuring reproducible and systematic probing of domain evolution. Our machine learning framework is embedded into the acquisition pipeline and performs on the fly extraction of polarization signatures from streaming 4D STEM data, while correlating these results with SEM and TEM observations to bridge length scales during the experiment itself. We introduce new approaches for classifying polarization domains and diffraction derived features across SEM and 4D STEM datasets using both supervised and unsupervised strategies. An unsupervised pipeline based on Disentangling Autoencoders [3] enables phase and domain differentiation directly from SEM data through clustering in latent space. In parallel, the latest iteration of the Ghost Hunter framework [4] employs a Group Convolutional Autoencoder model [5] to encode rotational equivariance and a Spatial Transformer network [6] to enforce translational invariance, enabling robust mapping of polarization vectors from diffraction features including ghost disks [4], Kikuchi bands [7], and intensity asymmetries [8]. The model is trained on large, simulated datasets and generalizes across multiple ferroelectric systems including PbTiO3, BaTiO3, and oxyhalide boracites under typical experimental conditions. A critical aspect of this work is quantifying how specimen misalignment and thickness variations influence polarization sensitive diffraction features. We developed a simulation framework spanning multiple thicknesses and a range of tilts in orthogonal directions to systematically evaluate the evolution and degradation of polarization contrast. In figures 1 and 2 we show that tracking integrated intensities within regions of interest around polarization relevant diffraction features, we establish quantitative limits beyond which polarization becomes indistinguishable. These results define practical alignment tolerances and inform deployment boundaries for real time machine learning analysis under realistic imaging conditions. Together, these developments enable automated and agent driven microscopy for functional ferroelectric device characterization. By coupling automated in situ biasing acquisition with live machine learning analysis at the microscope, we close the feedback loop between data collection, interpretation, and adaptive control. Applying this integrated framework across SEM and TEM modalities allows polarization dynamics to be tracked coherently from the atomic scale to mesoscale domain structures as experiments unfold. This approach advances ferroelectric device studies from static post-processing toward intelligent, real time, multiscale decision-driven microscopy [9]. Intensity comparison matrix for different tilt angles. (a) Definition of the two boxed regions of interest (ROIs) used for the analysis. (b) Corresponding asymmetry score matrix, calculated as the intensity difference between the two ROIs shown in (a); values > 0 indicate that the net intensity in ROI 1 exceeds that in ROI 2 and that the method is indicating the correct direction of polarization. a) CBED image of PbTiO3 showing the masked regions for the Kikuchi band COM analysis. b) COM matrix of masked boxes for simulated PbTiO3 at different tilt angles. Positive values of y COM indicates that the tilt angle is aligned with the true direction of polarization, negative shows that it is no longer indicating the correct direction of polarization and hence the method becomes inaccurate.