Deconvolution of electrochemical impedance spectroscopy data using the deep-neural-network-enhanced distribution of relaxation times

Abstract

Electrochemical impedance spectroscopy (EIS) is widely used to characterize electrochemical systems. The distribution of relaxation times (DRT) has emerged as a powerful, non-parametric alternative to circumvent the inherent challenges of EIS analysis through equivalent circuits or physical models. Recently, deep neural networks have been developed to estimate the DRT. However, this line of research is still in its infancy, and several issues remain unresolved, including the long training time and unknown accuracy of this method. Furthermore, deep neural networks have not been used for deconvolving DRTs with negative peaks. This work addresses these challenges. A pretraining step is included to decrease the computation time; error analysis allows error estimation and the development of error reduction strategies. Furthermore, the training loss function is modified to handle DRTs with negative peaks. For most cases tested, this new framework outperforms ridge regression. Moreover, these advances are validated with an array of synthetic and real EIS spectra from various applications, including lithium-metal batteries, solid oxide fuel cells, and proton exchange membrane fuel cells. Overall, this research opens new avenues for the development and application of the deep-neural-network-based analysis of EIS data.