I think that training on MD trajectories may be a mistake. Basically they are just creating a deep learning encoder that learns the MD weights, which are often notoriously inaccurate.
I can't actually tell what MD techniques they are using. Citation 28, which says they're described in supplemental materials, which I cannot find on the arxiv page.
Either way, I think the right way to do this is to run explicit ab initio quantum MD (IE, as rigorous as it gets), then train on those trajectories. They aren't weighted, like classical approximation force fields like AMBER or MP2 are.
The hard part will be crafting models that can handle any molecular topology based on a fixed set of learn-able parameters. What dataset will be used to train the model is up to each practitioner (i.e. PDB datbank, ab initio, ect).
I may be reading this wrong, though.