First-principles molecular simulations provide fundamental insights into the structure-property relationships of materials, but their high computational cost limits large-scale applications. With the advent of machine learning interatomic potentials (MLIPs), it has reduced computational costs by enabling large-scale dynamic simulations with accuracy comparable to density functional theory (DFT), thereby accelerating high-throughput screening in materials discovery. In this work, DysNet (dynamic and spherical network) is introduced─a novel SE (3)-equivariant graph neural network that addresses the tradeoff between physical fidelity and computational efficiency. Its core innovation is a chemically gated interorder attention (CG-IOA) mechanism, which learns to adaptively allocate contributions from different physical interaction orders. This mechanism is supported by two synergistic components: a computationally efficient, implicitly many-body message passing framework, and an initialized spherical harmonic tensor embedding that provides rich geometric priors. Across four diverse benchmark data sets─QM9, rMD17,3BPA and SPICE─the model demonstrates comparable or superior predictive accuracy to most state-of-the-art equivariant graph neural networks and demonstrates remarkable adaptability across diverse chemical benchmarks.
Wang et al. (Tue,) studied this question.