Analysis of Thousands of Electrochemical Impedance Spectra of Lithium-Ion Cells through a Machine Learning Inverse Model

Electrochemical impedance spectra of lithium-ion cells can be collected periodically at various cycle numbers and various state of charges, producing vast amounts of data. Fitting each spectrum to an equivalent circuit can lead to physical insights about the evolution of the lithium-ion cell, yet the fitting problem requires good human initial guesses for the circuit parameters to reliably converge, making the fitting process labor intensive and difficult to scale. This article presents a paradigm to automate the fitting of measured data to physical models, replacing the good human first guesses with an inverse model parametrized with an artificial neural network. This method is simple to implement, uses principles applicable to a wide variety of fitting problems, and leads to reliable and accurate initial guesses of the circuit parameters for a given spectrum. The software implementation will be freely available once a good user interface is developed, and the performance of the system is evaluated on a dataset of about 100000 impedance spectra from lithium-ion cells, achieving a failure of fitting approximately 1% of the dataset, corresponding to the percentage of poor quality data in the dataset.