The following dissertation discusses the usage of data-driven approaches, combined with density functional theory (DFT), to study the complex reactions involved in the battery systems, e.g., electrolyte degradation and solid electrolyte interphase (SEI) formation. Data-driven approaches include the development of machine learning (ML) interatomic potentials for the challenging charged/open-shell molecular species, reaction network construction based on high-throughput DFT, ML, automated fragmentation and recombination strategies to yield a thermodynamic landscape of millions of possible reactions for the SEI formation, and computational screening of novel electrolytes for multivalent ion batteries based on their kinetic barriers to reductive decomposition.Chapter 1 provides a brief overview of battery chemistry and the computational modeling tools for investigating such systems, e.g., ab-initio methods or classical/ML force fields. Furthermore, this overview highlights the challenges and the strategies for the computational modeling of battery materials, bulk electrolyte properties, and electrochemical and interfacial reactions.
Chapter 2 describes the development of the Becke Population Neural Network (BpopNN) model as an ML interatomic potential that is capable of handling charged molecules, which are essential in electrochemistry and battery systems. This is accomplished by incorporating electronic information (electronic population on each atom) into the descriptor and minimizing the total energy self-consistently with the total population constrained to be the total number of electrons in the system.
Chapter 3 extends from the work discussed in Chapter 2 and demonstrates the application of the BpopNN model to Li-organic systems, which have direct relevance to battery electrolytes. Spin populations are incorporated into the model, and possibilities of using BpopNN to model charge-transfer excited states are discussed.
Chapter 4 describes the study of using an automated reaction network and DFT to probe the formation mechanisms of Lithium Ethylene Monocarbonate (LEMC), a controversial organic component of the SEI in lithium-ion batteries (LIB). It manifests another application of data-driven methods and ML models to studying the complex reaction cascades or mechanisms in battery systems. It was found that water is essential in the formation of LEMC; therefore, LEMC is not likely to be the major component of the SEI if rigorously dried electrolytes are used. Several novel mechanisms for forming LEMC were also proposed in an automated fashion through a path-finding algorithm applied to the massive reaction network.
Chapter 5 describes a computational screening effort for novel electrolytes in the weakly-coordinating fluorinated alkoxyborate and alkoxyaluminate salt family for Mg/Ca-ion battery applications. The design metric employed is based on the redox potentials and the kinetics for reductive decomposition, covering various decomposition patterns, e.g., the breaking of Al/B-O, C-O, or C-F bonds. The kinetic barriers for such decomposition pathways are obtained and demonstrated to correlate well with their electrochemical performance for existing electrolytes. Novel electrolytes that are never synthesized before are also proposed based on this criterion.
Chapter 6 summarizes and concludes the thesis.
