Atomic-Scale Insights into Polarization Mechanism in HfO₂ Ferroelectrics Using Multi-Phase Machine Learning Potential

Recently, HfO₂ ferroelectrics have emerged as key materials for future semiconductor technologies. However, they face practical challenges due to high coercive fields and slow domain wall mobility, which impact their operational performance. To address these issues, it is essential to understand the phenomena associated with the polarization of HfO₂ ferroelectrics at the atomic level. Additionally, since HfO₂ has various metastable and stable phases, developing an accurate machine learning potential (MLP) that can simulate the energy stability between these multiple phases is important to reveal the polarization mechanisms at the atomic level. Therefore, in this study, we develop a multi-phase machine learning potential (MP-MLP) that accurately predicts the energy and force of various metastable and stable phases of HfO₂ with quantum mechanical precision, using a graph neural network-based model. The energy order of different phases predicted by the MP-MLP matches exactly with density functional theory (DFT) calculations. We also observe polarization rotation and switching in HfO₂ ferroelectricity using MP-MLP based molecular dynamics simulations under external electric field conditions and identify intermediate phases during polarization switching and rotation. These results provide atomic-level insights related to the wake-up and fatigue of HfO₂ ferroelectric devices. We expect these insights to facilitate the development of more efficient semiconductor devices, enhancing the performance and reliability of future technologies.