This paper proposes a new data-driven approach for modeling detailed splashes for liquid simulations with neural networks. Our model learns to generate small-scale splash detail for fluid-implicit-particle methods using training data acquired from physically accurate, high-resolution simulations. We use neural networks to model the regression of splash formation using a classifier together with a velocity modification term. More specifically, we employ a heteroscedastic model for the velocity updates. Our simulation results demonstrate that our model significantly improves visual fidelity with a large amount of realistic droplet formation and yields splash detail much more efficiently than finer discretizations. We show this for two different spatial scales and simulation setups.