Machine learning assisted classification and interpretation of EIS data with experimental demonstration for chemical conversion coatings on Mg alloys

Magnesium and its alloys, promising for structural applications, have significant corrosion concerns. Chemical conversion coating (CCC) is a cost-effective, scalable, and environment-friendly method to address this issue. EIS is a powerful technique frequently used in various applications, including CCCs, to model diverse electrode-electrolyte interface processes through equivalent circuits (EC). The traditional EC approach is tedious, requires human experience and imparts subjectivity. Machine Learning (ML) strategies automate the process of EC classification, identify complex patterns, reduce subjectivity and analysis time. This study focuses on the ML classification of ECs in CCCs, covering almost all realistic coating conditions and processes. We have used both linear (PCA: Principal Component Analysis and LDA: Linear Discriminant Analysis for interpretation and dimensional reductions; Logistic Regression for classification) and non-linear ML models (Naïve Bayes, Random Forest, XG Boost, Gradient Boost and 1D-CNN: all for seven-class EC classification). We observed that PCA reduced 213-dimensional feature space to 8 PCs, retaining 99% variance, whereas ∼89% of the top 10 features in LDA belong to low-frequency regions. The performance of ML models was evaluated by diverse metrics: confusion matrix, F1-score, precision, recall, Shapley Additive Explanations (SHAP), and Area under Receiver Operating Characteristic Curve (AUC). Overall, 1D-CNN model performed the best with an accuracy ∼86%, mean AUC ∼0.98, and top-2 accuracy ∼96%. From SHAP analysis, we found low-frequency phase measurements are critical for the 1D-CNN model's decision-making, signifying the importance of slow electrochemical processes. Finally, we validated best ML model with experimental data and found 1D-CNN classifier has an average R²∼0.96.