Quantitative analyses of DPS and PEG-PPG in Cu electrolyte using machine learning with artificial neural network

Abstract

The measurement of additive concentrations in Cu electrolyte is the first step in maintaining the performance of Cu electrodeposition. The cyclic voltammetric stripping (CVS) method has been continuously used for this purpose, despite its limitations, such as low accuracy and the inability to perform in-situ monitoring. Consequently, long-term usage of Cu electrolyte carries a risk of process errors due to fluctuations in additive concentration. To address these issues, this study introduces a machine learning (ML)-based technique to extract additive concentration information from a single voltammogram, without the need for any pretreatment or sampling steps. Specifically, this ML-based technique aims to predict the concentration of 3-N,N-dimethylaminodithiocarbamoyl-1-propanesulfonic acid (DPS), an accelerator that exhibits non-linear acceleration behavior depending on its concentration. Four different algorithms—linear regression, ridge regression, random forest, and neural network models—are examined for their ability to learn the complex interaction between polyether and DPS, allowing extraction of their concentrations from a single voltammogram. This study demonstrates that neural networks are the most effective for capturing non-linear patterns in voltammograms. Additionally, our results indicate that careful selection of the potential range for training can yield an efficient ML technique by minimizing model size while maintaining high analytical accuracy.