Production, characterization and prediction of battery performance using machine learning of silicon-based anode thin films for lithium-ion batteries

This thesis aims to improve the performance of silicon (Si)-based thin film anodes for lithium-ion batteries (LIBs) and consists of two main stages. In the first stage, discharge capacity is modelled and predicted using data science and machine learning (ML) techniques. Missing values in a multi-source dataset compiled from the literature were imputed using k-Nearest Neighbors (KNN), Gaussian Process Regression (GPR), Multiple Imputation by Chained Equations (MICE), and Random Forest methods; the Random Forest-based approach was selected as it provided the highest accuracy. The same model was used to extend capacity curves between the 5th and 500th cycles, generating physically reasonable capacity fading trends. To investigate how fabrication and electrochemical operating parameters affect discharge capacity and its decay with cycling, a model-agnostic Accumulated Local Effects (ALE) analysis was applied, from which practical "design windows" were proposed for next-generation Si thin film anode design. In the experimental stage, Si/MO/Si thin film anodes were fabricated on Cu current collectors by radio frequency (RF) magnetron sputtering without using any binder or conductive additive. Tungsten trioxide (WO₃), vanadium pentoxide (V₂O₅), tantalum pentoxide (Ta₂O₅), and molybdenum trioxide (MoO₃) were employed as metal oxide (MO) interlayers, and deposition parameters were optimized for nanometer-scale thicknesses. Using these interlayered architectures, the enhancement of interfacial stability and the mitigation of volume expansion-induced degradation in Si-based thin film anodes were targeted and evaluated based on electrochemical performance results.