Accurate prediction of the remaining useful life (RUL) of li-ion batteries (LIBs) is essential for enhancing the operational efficiency and safety of LIB-powered applications. It also facilitates improvements in the cell design process and the evolution of fast charging methodologies, thereby minimizing cycle testing time. While artificial neural networks (ANNs) have emerged as promising tools for this task, identifying the optimal architecture across diverse datasets and optimization strategies is non-trivial. To address this challenge, a machine learning framework was developed for a systematic evaluation of diverse ANN architectures. Utilizing just 30% of the training dataset from 124 li-ion batteries cycled under various charging policies, hyperparameter optimization is conducted within this framework. This ensures that each model is evaluated at its optimal configuration, facilitating a balanced comparison for RUL prediction tasks. Furthermore, the study examines the influence of varied cycle windows on model efficacy. Employing a stratified partitioning method highlights the importance of uniform dataset representation across different subsets. Notably, the top-performing model, using cycle-by-cycle features from just 40 cycles, achieves a mean absolute percentage error of 10.7%.