Molecular dynamics (MD) simulations are very useful to study how system of atoms evolve over time, but they are very slow and computationally expensive. Generative models can be used to learn to propagate molecular configurations forward in time, effectively replacing explicit numerical integration of newton's equation of motion, promising a dramatic speedup. However, to train these models, large volumes of trajectory data generated by MD simulations are required. Generating that training data remains a dominant cost. The project aims at finding the strategies to collect data for training transferable generative models. Once a model is trained, active learning methods can be used to expand the training regime of the model to improve it's performance. Specifically, we want to test if we can come up with a strategy to select a subset of the training data or select data in a prioritized manner to train the model which will improve the overall test performance of model faster than random subset selection. Such strategy will allow us to get understanding about what kind of similarities in the data the model is learning, thus giving us more insights about the factors in dataset that can affect the model's performance or introduce certain biases. It will also allow us to strategically simulate the more informative structures for the active learning tasks, saving hours of compute.