Enhancing the study of battery degradation using machine learning to get more out of data

Abstract

Improving energy storage technology is one of the key challenges for tackling anthropogenic climate change. This work investigates the degradation of two key technologies for energy storage. Lithium-ion batteries (LIBs) are the most promising candidate for use in electric vehicles and portable electronics, whilst redox flow batteries (RFBs) are a promising candidate for grid-level energy storage. Herein, the degradation of both LIBs and RFBs is investigated using machine learning to both forecast battery degradation and gain insight into the degradation mechanisms themselves. The first chapter investigates extracting features from electrochemical cycling data, in particular the charge and discharge curves, and using them to diagnose and forecast battery degradation. It is shown that these features can improve the accuracy of models forecasting battery health. A key output of this work is the Navani Python package, which processes and standardises electrochemical data from a wide variety of cyclers, enabling easy and consistent comparisons across different machines. The second chapter examines the use of convolutional neural networks (CNNs) to automatically analyse powder X-ray diffraction (PXRD) data and extract structural information from diffraction patterns in real time. This approach is validated on a real dataset of PXRD patterns collected on a cycling LiNixMnyCo(1-x-y)O2 (NMC) cathode. Lattice parameters are accurately extracted, and the heterogeneity of the electrode is measured in a novel manner. In some cases, the CNN outperforms traditional Rietveld refinement on the real PXRD data. In the third chapter, non-negative matrix factorisation (NNMF) is used to model in situ UV-Vis data from an RFB. The models can track the different molecules in solution during cycling and are able to detect both degradation products that had previously been observed, and the dimerisation of the 2,6-dihydroxyanthraquinone molecule, that had not previously been seen.