Convolutional Kernel-Embedded E (3)-Equivariant Networks (KEN): Overcoming Architectural Rigidity in Machine Learning Interatomic Potentials

Abstract

Abstract Machine learning interatomic potentials (MLIPs) have emerged as scalable alternatives to first-principles methods such as density functional theory (DFT). Among them, E(3)-equivariant graph neural networks (GNNs) like PACE and MACE are highly accurate but structurally rigid, limiting the incorporation of rich, multichannel electronic descriptors. We introduce convolutional kernel-embedded E(3)-equivariant networks (KEN), a hybrid architecture that combines the local environmental extraction of 2D convolutional kernels with the many-body expressivity of an E(3)-equivariant GNN backbone. This design preserves symmetry while enabling flexible, multimodal input features without increasing training data requirements. Trained and tested on heavy element 26 (HE26) dataset, KEN was compared with MACE-osaka26 model. Further validation on lattice parameter prediction through Birch–Murnaghan equation of state, magnetic moment prediction, and comparison with CHGNet framework, and microcanonical molecular dynamics confirms robust energy conservation and accurate reproduction of the forces, and thus, demonstrates a viable pathway to overcoming architectural rigidity in state-of-the-art MLIPs.