Towards Efficient Computational Predictions of Battery Cathodes

Abstract

A series of computational tools were employed with the primary goal of understanding layered transition metal oxide materials as lithium ion battery cathodes. As nearly all of the data within this thesis was generated from density functional theory (DFT), we begin with an analysis of the uncertainty of DFT with respect to the choice of exchange correlation functional through the development of a prediction confidence metric and the propagation of error through the Debye-Gruneisen model for lattice vibrations. The prediction confidence metric is applied to the study of transition metal ordering in layered Ni-Mn-Co (NMC) oxide cathodes and enables us to rationalize the disagreement with experimentally seen phases. For the purpose of accelerating the computational predictions of these materials, we then train a neural network potential for the prediction of energy and forces using atom centered symmetry functions as the featurization. The success of this highly accurate machine learning potential is seen through its ability to recreate the thermodynamic properties with an added error that is below the error of the underlying DFT itself. We then predict the open circuit voltage for a series of NMC compositions as well as the lattice dynamics during cycling that have been linked to degradation of the cathode. We then quickly explore a promising machine learning algorithm that is beyond the fingerprint based methods conventionally used. Finally, we dive deeper into the mechanism of another avenue of degradation in the release of highly reactive singlet oxygen seen in NMC, as well as Li-air and Na-air batteries. We provide a unified picture for the mechanism, effect of electrolyte properties, and onset potential for this singlet oxygen release.