What is the appropriate data representation of electrochemical impedance spectroscopy in machine-learning analysis?

Electrochemical impedance spectroscopy (EIS) is an important analytic technique for the understanding of electrochemical systems. With the recent advent and burgeoning deployment of machine learning (ML) in EIS analysis, a critical yet hitherto unanswered question emerges: what is the appropriate data representation of EIS for ML-based analysis? While the representation of a model’s input data is known to be critical for a successful deployment of ML model, EIS is known to possess multiple classical venues of data representation and it remains unclear how different EIS data should be compared following a proper data normalization protocol. Here we report the methodology and the outcomes that evaluate the efficacy of multiple data representation methods in ML-based EIS analysis. At least within our proof-of-concept parameter space, plotting the input training data’s impedance magnitude (|Z|) against phase angle (φ) while individually normalizing each EIS curve yields the highest accuracy and robustness in the correspondingly established residual neural network (ResNet) model. Rationalized by additional "importance" analysis of the input data, such a data representation method extracts information and hidden features more effectively. While Nyquist plot is more widely used in manual analysis, we found that ML-based analysis may require a different data representation and offered a clear guideline for future researchers to evaluate on a case-by-case basis.