Predictive model for SiO2 film properties using plasma optical emission spectra based on machine learning

Optical emission spectroscopy (OES) is widely used in the semiconductor industry for monitoring and controlling the SiO 2 deposition process without interfering the plasma. However, it is quite challenging for tetraethylorthosilicate (TEOS)-based plasma due to its complex emission spectrum. Furthermore, oxygen atom density and oxygen ions flux become important parameters when oxygen is presence in the deposition process. In this study, we propose an approach that expresses the SiO 2 deposition rate using a mathematical model based on physical mechanisms by leveraging plasma emission spectrum data and machine learning. A capacitively coupled plasma generated by RF power of 13.56 MHz was employed to deposit SiO 2 films with varying plasma power, working gas pressure, and gas ratio of TEOS/oxygen/argon. Furthermore, a gradient boosting regression tree model was used to predict the deposition rate using plasma emission spectrum, and then model prediction result was then interpreted using SHapley Additive exPlanation (SHAP). The SHAP result revealed that the deposition rate is significantly influenced by OH (308.95 nm), CO (560.84 nm), OI (844.72 nm), ArI (420.08 nm), and ArII (487.89 nm). It is suggested that these species are correlated to the TEOS dissociation process through electron impact, reaction with oxygen atoms and argon metastable. Furthermore, we used this result to derive a mathematical model for determining the SiO 2 deposition rate, indicating that OES could be used to monitor the plasma output under variation process parameters. • Machine learning application for plasma process using optical emission spectroscopy • SiO 2 deposition depends on the following plasma species CO, OH, OI, ArI, and ArII. • SiO 2 deposition rate is possible to calculate from plasma optical emission spectrum. • Explainable machine learning provides insight on the SiO 2 deposition by PECVD.