Modeling complex system dynamics traditionally is implemented with the use of differential equations, which requires hand-crafted work of a qualified expert and significant amount of time. The advent of data-driven approaches allows to overcome these difficulties and substitute traditional models with models built in automated way directly from observations. This paper compares several data-driven approaches to modeling 2D liquid simulator. Dataset is generated from it for both training and testing with fixed simulator parameters. Local and global types of models are evaluated with metrics, describing different aspects of liquid behavior (spatial, spatio-temporal and worst-case settings). Other metrics introduced allow to capture differences not only in distances, but also in distributions, which is more natural for human perception and enables to quantitively compare similar pictures. From the model evaluation, it is inferred that the use of decomposition improves overall accuracy and the trajectories figures, though at the same time model generalizability decreases. On the other hand, utilizing locality leads to more generalizable models at the cost of accuracy. Model training and inference times are provided and main directions for further research are outlined.