Spin-Weighted Spherical Harmonics Enable Complete and Scalable $\mathrm{E}(3)$-Equivariant Networks

Researchers have developed a method using spin-weighted spherical harmonics to perform geometric deep learning more efficiently. By replacing the computationally expensive Clebsch-Gordan tensor product with a new Gaunt tensor product, this approach lowers the complexity of E(3)-equivariant networks. This improvement allows for more scalable modeling of complex 3D structures, such as large atomistic systems, which were previously limited by high processing requirements.