Neural network approach for ferroelectric hafnium oxide phase identification at the atomistic scale

The hafnia-based ferroelectric oxides with excellent negative-capacitance properties offer a great opportunity to develop high-performance integrated circuits. The nanosized multiphase distribution of Hf_0.5Zr_0.5O₂ (HZO) has a significant influence on its ferroelectric properties. Transmission electron microscope (TEM) with an atomistic resolution could establish the structure-property relationship and guide the performance improvement of HZO by identifying its phase structures. However, the high throughput TEM data and its complexity of interpretation make the quantitatively extracting the physical and chemical information from the TEM images challenging and low-efficiency. Here, we develop an automatic work flow for the TEM data analysis, which greatly enhances the efficiency of TEM data processing. By extracting the interest area and training the neural network with ResNet18, the accuracy of phase determination reaches 95.82% with low computational cost. Theoretical analysis is conducted to unveil the advantages of the ResNet18 network. The approach provided in this work could promote the quantitative analysis of the high-throughput TEM images and pave the way for future on-line analysis of the TEM image stream in real-time.