Machine learning for quality control of Tin-Copper electrodes

• Machine learning ensemble model classifies 100 electrodes in under 100 ms. • Random forest and support vector machines reach 92% accuracy in manufacturing quality control. • Overfitting reduced with Synthetic Minority Over-sampling Technique. Screen printing is the most common technique for manufacturing printed electrodes, yet machine learning applications in its quality control are underexplored. Herein, SnO 2 -CuO microparticle-nanoparticle-based electrodes were printed on paper substrates and categorised as bad, medium, or good. The optimal chemical composition of the formulated inks and the maximum operating temperatures of the electrodes (up to 150 °C) were investigated. The inks were formulated using various ratios of SnO 2 powder (particle size 325 µm), glycerine and Polyvinyl alcohol (PVA) glue. The highest conductivity recorded was 80 S.m −1 achieved for an ink composed of 1 g SnO 2 powder, 3 g glycerine, 0.03 PVA glue, and heat treated at 150 ⁰C for 30 min. Random forest (RF), support vector machine (SVM) and Feedforward neural networks (FNN) models achieved approximately 80 % accuracy and recall. Synthetic Minority Over-sampling Technique with Edited Nearest Neighbors and hyperparameter tunning methods were used to eliminate model overfitting by balancing the dataset. An ensemble model that can classify about 100 electrodes in under 100 ms and can correctly classify bad electrodes with an average accuracy of 92 % was developed. This ensemble model was demonstrated as a feasible solution for real world fast-paced manufacturing environments.