Do we need all the parities for the hidden layer features ?

[MSJavaScript]

Hello, I am trying to use nequip to build my own model and I find that in Nequip, both odd and even irreducible tensors are used for the hidden layer. However, in mace, only the physical parity is used. For example, in mace, the hidden layer features may be 64x0e + 64x1o, but in nequip, it will be 64x0e+64x0o+64x1e+64x1o, which one is reasonable ？

[cw-tan]

In NequIP, you can switch it with the parity argument -- see API: https://nequip.readthedocs.io/en/latest/api/nequip_model.html#nequip.model.NequIPGNNModel . My advice would be to try both and see if you can get away with parity: false, if the accuracies are similar.