Abstract
Electrochemical impedance spectroscopy (EIS) is entering an exciting stage of development as machine learning–driven (ML) spectral analysis begins to complement traditional equivalent-circuit fitting and address some of its practical limitations. The appeal of a simple, fully data-driven, unsupervised workflow is clear: by operating directly on EIS spectra, it bypasses the additional modelling layers required for equivalent-circuit fitting, handcrafted feature extraction, or supervised training. Here we demonstrate that normalization and dimensionality reduction play a critical, yet previously overlooked, role in shaping the outcomes of unsupervised workflows. Using welded stainless steel as a demonstrator, we systematically evaluate combinations of normalization strategies and dimensionality-reduction pipelines. By applying internal clustering metrics and a Borda ranking, we identify an effective workflow configuration, an appropriate cluster number, and a cluster structure consistent with mechanistic expectations for the studied dataset. Mechanistically anchored linear projections further rank relative passivity across the stainless-steel passivity range via k-level clustering, while bootstrap resampling confirms high cluster stability despite the modest sample size.