Equivalent Electrical Circuit recommendation for Electrochemical Impedance Spectroscopy: A benchmark of different Machine Learning algorithms

• Different ML algorithms were benchmarked for EEC recommendation for EIS. • The hyperparameters of each algorithm were optimized. • The accuracy distributions were obtained (instead of point estimates). • The performance of the different recommendation algorithms was compared. • The best algorithm was used to recommend an EEC for real EIS spectra. Electrochemical Impedance Spectroscopy (EIS) is a technique widely used in the electrochemistry field that allows assessing the kinetic parameters of electrochemical systems. In general, EIS spectra are analyzed using 2 methodologies: the physical–mathematical methodology and the Equivalent Electrical Circuit (EEC) methodology. The physical–mathematical methodology consists in developing and solving the set of differential equations that govern the system. In contrast, the EEC methodology describes EIS data in terms of EEC models, decreasing the math complexity and democratizing the use of the EIS technique. In the EEC methodology, the most critical task is selecting a physically sound EEC model. To help with the EEC model selection, Prof. Digby D. Macdonald proposed creating the Electrochemistry’s Genome Project, which would consist of a large database and an Artificial Intelligence (AI) able to suggest plausible EEC models given an EIS spectrum. Until now, several works have explored the idea of using Machine Learning (ML) for recommending an EEC for a given EIS spectrum. Indeed, several authors have benchmarked different ML algorithms for this task. However, these benchmarks generally report point estimates of the accuracy achieved by each algorithm, but do not assess its uncertainty. In this work, we benchmarked different ML EEC recommendation algorithms. To achieve this, we refined a preexistent database, obtaining a new database. Then, the new database was used to train different ML algorithms, optimize their hyperparameters, and assess their accuracy. Finally, the performance of the different optimized algorithms were compared. Whereas most of the other benchmarks available in literature work with point estimates of the accuracies, in this work, the accuracy distributions have been estimated. This methodology allows to compare the performance of the different algorithms in a much more reliable and robust way.