Artificial intelligence for electrode quality control in battery manufacturing: Evaluating machine learning, deep learning, and transfer learning models for Tin, Zinc, and Titanium electrodes

Screen printing enables high-throughput, high resolution, and low-cost battery manufacturing, but manual inspection cannot keep up with production rates. Building on three prior single-material quality control classification publications (ZnO, SnO 2 , TiO 2 ) that each exceeded 98% quality-grading accuracy trained on over 900 electrode images, this follow-up article asks the harder, industrially realistic question: can one model simultaneously identify material and quality across seven classes? Herein, four machine-learning models (Random Forest, XGBoost, Support vector machines (SVM), Logistic Regression), two deep-learning models (Feed forward neural networks, Convolutional neural networks), and one transfer-learning model (ResNet-50) were investigated. To make classical ML viable for images, we introduce a compact image-to-feature pipeline that converts electrode photos into robust descriptors, enabling fast training/inference without heavy compute. Macro-F1 is the primary metric to address class imbalance and asymmetric costs. ResNet-50 delivers the best overall performance (Accuracy = 0.92, Precision = 0.87, Recall = 0.90, F1 = 0.88) with low overfitting. SVM provides the most stable classical alternative with modest accuracy and minimal overfitting. Results show that models struggle when multiple materials are involved, and that transfer learning offers a practical route to real-time, multi-material quality control in battery-electrode manufacturing.