Real-time prediction of dielectric properties in a HZO-Based thin-film capacitor using deep learning

Abstract

The dielectric properties of hafnium–zirconium oxide (HZO)-based thin-film capacitors are critical for optimizing device performance and improving device reliability. Conventional methods for characterizing these properties rely on direct electrical measurements, which are time-consuming and unsuitable for large-scale semiconductor production. This study presents an AI-based method that rapidly predicts the dielectric properties of HZO thin films from microscopic image data. A convolutional neural network (CNN) model was trained to infer dielectric behavior based on color changes induced by post-metal annealing (PMA) process conditions. These surface color variations are directly linked to internal structural phase transitions, supporting the use of optical features as indicators of dielectric phase states. The model achieved prediction accuracies of 63 % and 50 % when trained on image data from HZO and Mo regions in HZO thin-film capacitors, respectively. Using combined image data from both regions improved accuracy to 88 %, highlighting the significance of capturing both HZO crystal structure changes and Mo electrode oxidation effects. This AI-based inspection technique enables rapid, non-contact classification of dielectric properties at the chip level before the packaging process, offering a practical tool for early detection and process optimization in semiconductor manufacturing. In addition, this AI-based imaging inspection may work as a sensor module for the real-time feedback control system to achieve the targeted dielectric properties. While demonstrated on HZO, the approach can be extended to other thin-film materials where physical or chemical changes affect surface appearance, providing a scalable and cost-effective platform for inline quality control across various device fabrication processes.