Optimization of a convolutional neural network for system identification in electrochemical impedance spectroscopy

Electrochemical Impedance Spectroscopy (EIS) is a well-stablished electrochemical technique. Commonly, EIS data is analyzed using the Equivalent Electrical Circuit (EEC) approach, which consists in developing a physically sound EEC that can fit the impedance data. The first step to propose an EEC is the identification of the system. In EIS, the first stage of the identification of the system consists in quantifying the number of time constants of the system. Commonly, the identification of the system is carried out using either the visual method or the Distribution of Relation Times (DRT) method. For the identification of an individual EIS spectrum, both methods can yield reliable results when performed by an expert. However, for large databases of EIS spectra, their application becomes time-consuming and impractical. In this work, a Convolutional Neural Network (CNN) was optimized for system identification of purely capacitive EIS spectra. The optimization was conducted using a two-stage Design of Experiments (DoE). The first stage optimized the convolutional network; and then, the second stage optimized the dense network. At the end of the optimization process, an optimized CNN with an overall test accuracy of 92.5% (above 98% for spectra containing 1 or 2 time constants; and above 80% for spectra containing 3 or 4 time constants) was obtained. The algorithm developed in this work is a math-free (from the user's perspective) and reproducible alternative to both, visual and DRT methods, for system identification in EIS. Besides, when applied to large databases of EIS spectra, the new algorithm can be easily automated, whereas the traditional methods cannot. • A new database comprising 48,000 capacitive EIS spectra was generated. • The hyperparameters of a CNN were optimized using a two-stage DoE. • A model interpretability technique was used to get insights into learned patterns. • The best CNN achieves a test accuracy of 92.5% with minimal class confusion. • The best CNN is a fast and user-friendly tool for system identification.