equivariant-neural-networks

The paper introduces EquiFiLM, a lightweight extension that adds continuous external conditioning to equivariant foundation machine learning force fields via Feature-wise Linear Modulation, achieving significant accuracy improvements with minimal training data.

This paper introduces SpinGTP, a method using spin-weighted spherical harmonics to achieve complete and scalable E(3)-equivariant networks for 3D atomistic simulations, recovering antisymmetric interactions lost in prior Gaunt Tensor Product approaches.

MALOQ introduces a massively accelerated machine learning model for predicting density functional theory Hamiltonian/density matrices, enabling electronic-structure calculations for systems with up to 100k atoms using an SO(2)-equivariant backbone and scalable graph distribution, achieving over 30% time-per-epoch reduction on the Alps supercomputer.