Machine learning-based etch rate prediction model in plasma dielectric etch

Precise control of dielectric film thickness is essential in plasma etching process, as it is a critical factor in the performance and reliability of semiconductor devices. Conventional thickness measurements are performed using ex-situ methods, which result in time delays between processing and measurement. To address this limitation, we developed an etch rate prediction model based on in-situ sensor data. A machine learning (ML) regression model was developed using Optical Emission Spectra (OES) data collected in real-time during the etching of SiO₂ films with NF₃/H₂ gas chemistry. The process conditions were varied by substrate temperature, gas flow rate, and gas composition ratio. Since the ML model is based on sensor data reflecting plasma characteristics, understanding the effect of process condition-induced OES variations on the etch rate is necessary. Therefore, to interpret the correlation between input and output data in the ML model, an eXplainable AI (XAI) method was applied, enabling the identification of key wavelengths that influence the etch rate. The proposed data-driven framework analyzes the process mechanism by predicting the etch rate and interpreting the contributions of each plasma species, thus demonstrating the potential for real-time process feedback.