Multivariate Statistical Analysis of High-Resolution Hyperspectral Image Datasets Using Nanoscale Capacitance-Voltage Measurements for Ferroelectric Thin Film Characterization

The key to understanding various unexplained phenomena related to ferroelectric materials and discovering new ones, as well as improving the properties of these materials, is to comprehend the domain dynamics at the nanoscale. Among various microscopy techniques, piezoresponse force microscopy (PFM) has been widely used for this purpose.[1,2] Another approach is the recently developed local capacitance-voltage (C-V) mapping method, which allows nanoscale analysis of domain dynamics through dielectric rather than piezoelectric measurements.[3,4] This method has advantages over PFM-based methods, such as high-speed and high-sensitivity observation, enabling quick acquisition of high-resolution hyperspectral image data. Here, we present a methodology to analyze C-V datasets using a machine learning approach to extract and visualize the information necessary for understanding ferroelectric domain dynamics.<br/><br/>Using this methodology, we focus on the effect of grain boundaries on polarization switching in ferroelectric films, an issue that has been debated for a long time and is critically important. We chose doped HfO₂ thin films as the specific target of our measurement. This material has outstanding properties, such as maintaining ferroelectricity even at a thickness of less than 20 nm, and is expected to be used in applications such as nonvolatile memory and neuromorphic devices. However, to realize such devices, it is necessary to solve problems related to the change in properties during cyclic electric field application, called wake-up and fatigue. In this study, we analyzed doped HfO₂ films before wake-up using local C-V mapping. The acquired datasets were then clustered into several regions based on the similarity of their polarization properties using unsupervised learning methods such as k-means and Gaussian mixture models (GMMs). In addition to the typical butterfly-shaped C-V curves that represent normal polarization switching, these clusters contained asymmetric butterfly curves that may be caused by domain pinning or other built-in field-derived effects. Subsequent statistical analysis suggested that, in pristine HfO₂ films before wake-up, although defects at grain boundaries have some effect on the polarization switching properties, they are not the main cause of the variation in properties.<br/><br/>[1] S. Jesse, A.P. Baddorf, and S. V. Kalinin, Appl. Phys. Lett. 88, 062908 (2006).<br/>[2] N. Balke, I. Bdikin, S. V. Kalinin, and A.L. Kholkin, J. Am. Ceram. Soc. 92, 1629 (2009).<br/>[3] Y. Hiranaga, T. Mimura, T. Shimizu, H. Funakubo, and Y. Cho, J. Appl. Phys. 128, 244105 (2020).<br/>[4] Y. Hiranaga, Y. Noguchi, T. Mimura, T. Shimizu, H. Funakubo, and Y. Cho, ACS Appl. Nano Mater. 7, 8525 (2024).