Questions about FullyConnectedTensorProduct in the code

[QiaolinLu]

Dear community,

I have a question in the code. In the NequIP network architecture, there’s a component defined as:
(sc): FullyConnectedTensorProduct(16x0e x 3x0e -> 48x0e+16x1o+16x2e | 2304 paths | 2304 weights)

This suggests that the input consists of scalar features (16x0e and 3x0e), yet the output includes higher-order tensor features (48x0e, 16x1o, and 16x2e).

Is it physically meaningful for the tensor product of two scalar inputs to yield higher-order tensor outputs?
How does the FullyConnectedTensorProduct module facilitate this transformation?

[cw-tan]

(sc): FullyConnectedTensorProduct(16x0e x 3x0e -> 48x0e+16x1o+16x2e | 2304 paths | 2304 weights)

is just the code doing admin for accounting purposes. If I'm not mistaken, only the scalar -> scalar paths actually do something and the rest of the output is just padded with zeros, probably to get the output in the right shape for further operations.

[QiaolinLu]

Got it, really appreciate for your reply!