Hybrid Deep Learning Crystallographic Mapping of Polymorphic Phases in Polycrystalline Hf_0.5Zr_0.5O₂ Thin Films

Abstract

By controlling the configuration of polymorphic phases in high‐k Hf0.5Zr0.5O2 thin films, new functionalities such as persistent ferroelectricity at an extremely small scale can be exploited. To bolster the technological progress and fundamental understanding of phase stabilization (or transition) and switching behavior in the research area, efficient and reliable mapping of the crystal symmetry encompassing the whole scale of thin films is an urgent requisite. Atomic‐scale observation with electron microscopy can provide decisive information for discriminating structures with similar symmetries. However, it often demands multiple/multiscale analysis for cross‐validation with other techniques, such as X‐ray diffraction, due to the limited range of observation. Herein, an efficient and automated methodology for large‐scale mapping of the crystal symmetries in polycrystalline Hf0.5Zr0.5O2 thin films is developed using scanning probe‐based diffraction and a hybrid deep convolutional neural network at a 2 nm2 resolution. The results for the doped hafnia films are fully proven to be compatible with atomic structures revealed by microscopy imaging, not requiring intensive human input for interpretation.

An Aqueous Route to Oxygen-Deficient Wake-Up-Free La-Doped HfO2 Ferroelectrics for Negative Capacitance Field Effect Transistors

“…The high-angle annular dark field (HAADF) imaging mode with a detector inner angle of ∼68 mrad was used to obtain atomic structure images of the samples. To distinguish crystal structures, position-averaged convergent beam electron diffraction (PACBED) and fast Fourier transform (FFT) analyses were subsequently conducted on the same regions observed by HAADF STEM. , Further, the chemical compositions at various depths of amorphous and crystalline HLO thin films were analyzed using depth profile X-ray photoelectron spectroscopy (XPS: ESCA 2000, MultiLab) with a monochromatic Al Kα X-ray source. The XPS spectra at different depths were recorded after etching with Ar for 300, 600, 900, and 1200 s. The recorded spectra were calibrated using the C 1s signature response at 284.6 eV.…”

“…Notably, the films grown using the 2-ME+Aq-based precursor show a single-phase configuration, even at a thickness of ∼9 nm. Due to the structural similarities of the c-, t-, and ophases, 26,27 an extended investigation using atomic-resolution scanning transmission electron microscopy (STEM) combined with position-averaged convergent beam electron diffraction (PACBED) and energy dispersive X-ray spectroscopy (EDX) was utilized. 26,28,29 1c and Figure S8) reveal that La is uniformly substituted for the Hf host without segregation and that the doping concentration is ∼7% relative to the Hf.…”

“…Due to the structural similarities of the c-, t-, and ophases, 26,27 an extended investigation using atomic-resolution scanning transmission electron microscopy (STEM) combined with position-averaged convergent beam electron diffraction (PACBED) and energy dispersive X-ray spectroscopy (EDX) was utilized. 26,28,29 1c and Figure S8) reveal that La is uniformly substituted for the Hf host without segregation and that the doping concentration is ∼7% relative to the Hf. These observations, made from a relatively large field of view, demonstrate that the HLO films grew uniformly with a flat surface, regardless of the solvent.…”

Decoding Material Structures with Scanning Electron Diffraction Techniques

“…Xu, W. and LeBeau, J. M. demonstrated the capability of CNN to automatically analyze PACBED, incorporating parameters such as zero-order disk size, center position, rotation, and sample details like thickness and tilt [27]. Recently, Kim, Y. H. et al utilized a hybrid deep CNN, combining supervised and unsupervised components, to analyze SED-based PACBED patterns in polycrystalline Hf 0.5 Zn 0.5 O 2 films [28] (Figure 3c). Their study highlighted that deep learning can effectively extract crystallographic information from PACBED patterns, even though the patterns are averaged over a small area of only 2 nm 2 .…”

Reduced dimensional ferroelectric domains and their characterization techniques

“…To extract precise peaks, the signal-to-noise ratio is improved using the Wiener filter in the frequency domain, especially for periodic images [213] . Advanced techniques such as nonlocal means denoising [214] and block matching and three-dimensional filtering (BM3D) [215,216] have also been employed. Currently, a significant portion of image analysis through peak detection has been automated, and the processing speed has greatly improved, alleviating many challenges in the structural analysis of ferroelectrics [208] .…”