Classifying metal passivity from EIS using interpretable machine learning with minimal data

We present a data-efficient machine learning framework for diagnosing degradation of passive metallic surfaces using Electrochemical Impedance Spectroscopy (EIS). Passive metals such as stainless steels and titanium alloys rely on nanoscale oxide layers for corrosion resistance, critical in applications from implants to infrastructure. Ensuring their passivity is essential but remains difficult to assess without expert input. We develop an expert-free pipeline combining input normalization, Principal Component Analysis (PCA), and a k-nearest neighbors (k-NN) classifier trained on representative experimental EIS spectra for a small set of well-separated classes linked to distinct passivation states. The choice of preprocessing is critical: normalization followed by PCA enabled optimal class separation and confident predictions, whereas raw spectra with PCA or full-spectra inputs yielded low clustering scores and classification probabilities. To confirm robustness, we also tested a shallow neural network (NN) classifier on the same PCA-reduced input, which produced comparable but slightly less confident predictions than k-NN. The training dataset was comprised of five classes including abraded (ABR), cold-rolled (INT), mirror-polished (POL), citric (CPAS) and nitric acid (NPAS) passivated states on AISI 304 stainless steel. The external validation set included impedance spectra from distinct samples and literature-reported datasets spanning crevice zones, welds, mechanically abraded regions, as-received materials, and passive surfaces in low corrosivity media. The results supports the viability of interpretable, label-efficient classification of EIS spectra for corrosion diagnostics.