Machine learning-powered recognition of crystalline phases and orientations in epitaxial Y-doped HfO2 via atomic-resolution STEM

Abstract

This study presents an efficient method for automatically identifying the crystal structure and orientation of Y-doped HfO2-based thin films using deep learning. This approach enables large-scale crystallographic analysis with sub-nanometer spatial resolution using only scanning transmission electron microscopy (STEM) atomic images, thereby reducing the reliance on manual expert interpretation. The Xception network-based model extracts detailed crystallographic information through structure and entropy maps, effectively identifying subtle pattern changes and local structural discontinuities. Entropy maps are utilized to analyze the atomic structure disorder and detect ambiguous boundaries and strained regions. Analysis of Y-doped HfO2 thin films reveals that the film thickness significantly affects the ferroelectric properties, with the O phase dominant in 5 nm films and the M phase proportion increasing as the thickness increases. This machine-learning-based STEM atomic image analysis method provides an automated solution to accelerate ferroelectric material research and promote the development of next-generation electronic devices, offering an accurate understanding and control of microstructural characteristics.