Why is this cosine scaling term in the SchNet implementation?

[NiklasKappel]

This line applies a cosine-based smooth cutoff to the filter values learned by the SchNet model. I guess it makes sense to have, as it reduces the interaction strength between more distant nodes. However, I can not find a reference to it in the paper on which the implementation is based, so I wonder why it is there?

[rusty1s]

This is taken from here: https://github.com/atomistic-machine-learning/schnetpack/blob/2ed8d1a3b773f4ed2dbd50623d43d578ff0146f6/src/schnetpack/nn/cutoff.py#L14