EqFSA: SE(3)-Equivariant Fourier Space Attention

We aim to address the over-squashing problem of traditional message-passing neural networks (MPNNs) while improving representations of long-range graph interactions. To this end, we introduce Eq-FSA, an SE(3)-equivariant architecture that uses a bounded-radius local branch and a Fourier-space master-node channel that learns input-conditioned spectral weights and lifts invariant global context into equivariant atom features via Clebsch-Gordan tensor products. We evaluate Eq-FSA on graph-transfer tasks designed to isolate over-squashing and on a controlled long-range electrostatics benchmark against RSA. Initial results are promising as Eq-FSA-LR reduces force MAE by 37% over a fine-tuned short-range backbone alone and by 90% over an invariant-only long-range baseline, while exhibiting improved long-range information propagation compared with local equivariant MPNNs.