Artificial Neural Networks (ANNs) are increasingly used in Computational Fluid Dynamics (CFD) for surrogate modeling, real-time flow prediction, and digital twin applications. In this study, the performance of different ANN architectures trained on data obtained from CFD simulations of a backward-facing step geometry are investigated and compared. The aim is to evaluate prediction accuracy and computational cost across varying network depths and structural configurations. The results highlight the trade-offs between model complexity, training time, and prediction accuracy, and provide recommendations for ANN-based surrogate modeling in CFD applications.