Phase Mapping of Polycrystalline Hf0.5 Zr0.5 O2 Films Using 4D-STEM with Unsupervised Learning

Abstract

Hf₀.₅Zr₀.₅O₂ (HZO) exhibits processing-dependent polymorphism, forming multiple crystallographic phases that critically influence its electrical properties. However, conventional characterization methods face significant limitations: X-ray diffraction (XRD) suffers from overlapping peaks, while transmission electron microscopy (TEM) provides only localized information with limited statistical coverage. These challenges necessitate high-resolution, large-area analytical techniques capable of quantitative phase mapping. In this study, we developed a comprehensive phase-analysis workflow combining four-dimensional scanning transmission electron microscopy (4D-STEM) with unsupervised machine learning. Nanobeam electron diffraction (NBED) patterns were processed using non-negative matrix factorization (NMF) and k-means clustering to identify structurally distinct regions. Phase indexing was performed through automated d-spacing extraction and template matching, with validation using simulated diffraction patterns. Applied to HZO films on La₂/₃Sr₁/₃MnO₃ substrates, this workflow yielded phase fractions consistent with XRD while providing spatially resolved nanoscale phase distributions. The methodology was subsequently applied to investigate degradation mechanisms in HZO-based memristor devices under repeated electrical cycling. Large-area 4D-STEM phase mapping revealed progressive transformation from ferroelectric orthorhombic to non-ferroelectric monoclinic phase as devices transitioned through pristine, degraded, and breakdown states. Complementary HRTEM, grain- size analysis, and EELS measurements confirmed concurrent oxygen vacancy redistribution and microstructural reorganization, establishing a direct link between field-induced orthorhombic phase instability and device failure. This work demonstrates that integrating 4D-STEM with unsupervised learning enables rapid, statistically robust phase characterization, overcoming limitations of conventional techniques and providing critical insights for next-generation ferroelectric and resistive switching memory devices.