Extrapolability of machine learning for determining additive concentrations in copper electrolyte

Abstract

Measurement of additive concentrations in acidic electrolytes is essential for maintaining the high reliability of the Cu electrodeposition process. Conventional cyclic voltammetric stripping (CVS) methods limit both accuracy and efficiency, often requiring excessive analytical time, and are not inherently suitable for in-situ analysis. An advanced analytical tool is needed to improve the Cu electrodeposition process, with machine learning (ML)-based techniques emerging as leading candidates due to their ability to simultaneously extract the concentrations of multiple additives from a single voltammogram. While ML-based tools ensure higher accuracy and significantly shorter analytical times, further research is necessary to advance these techniques. This study focuses on the extrapolability of ML models for determining additive concentrations. A case study involving the PEG-Cl^--SPS combination is conducted by intentionally limiting the concentration range for model training, followed by prediction of the three additives across the full concentration range. Various ML models are evaluated in terms of both upward and downward extrapolability by investigating the relationship between test and predicted concentrations of the three additives. The artificial neural network (ANN) model demonstrates the highest extrapolability in simultaneously determining the concentrations of three additives.