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

Area of Science:

• Materials Science
• Corrosion Engineering
• Machine Learning Applications

Background:

• Passive metals like stainless steel and titanium rely on oxide layers for corrosion resistance.
• Assessing the integrity of these passive layers is vital for critical applications but often requires expert analysis.
• Existing methods for evaluating passive surface states can be complex and time-consuming.

Purpose of the Study:

• To create a data-efficient machine learning framework for diagnosing degradation of passive metallic surfaces.
• To develop an expert-free pipeline for analyzing Electrochemical Impedance Spectroscopy (EIS) data.
• To enable reliable classification of distinct passivation states using limited labeled data.

Main Methods:

• Developed an expert-free pipeline combining input normalization, Principal Component Analysis (PCA), and a k-nearest neighbors (k-NN) classifier.
• Trained the classifier on experimental EIS spectra representing five distinct passivation states of AISI 304 stainless steel.
• Validated the framework using external datasets including literature-reported spectra from various conditions and environments.

Main Results:

• Normalization followed by PCA preprocessing significantly improved class separation and prediction confidence compared to raw spectra or other methods.
• The k-NN classifier achieved confident predictions, outperforming a shallow neural network (NN) on the same PCA-reduced input.
• The framework demonstrated robustness across diverse validation datasets, including literature data and various material conditions.

Conclusions:

• The developed machine learning framework offers a viable, label-efficient approach for corrosion diagnostics using EIS.
• Interpretable classification of EIS spectra is achievable, reducing reliance on expert interpretation.
• This method enhances the ability to ensure the integrity of passive metallic surfaces in demanding applications.