ViSNet is a geometric deep learning model designed to enhance molecular modeling by efficiently extracting geometric features and modeling molecular structures with low computational costs. It outperforms state-of-the-art methods on MD benchmarks like MD17, revised MD17, and MD22, and achieves excellent performance on QM9 and Molecule3D datasets. ViSNet efficiently explores conformational space and provides interpretable geometric-to-structural mappings. It uses an equivariant graph neural network (EGNN) approach, incorporating runtime geometry calculation (RGC) and a vector-scalar interactive message passing (ViS-MP) mechanism. RGC implicitly extracts geometric features like angles and dihedral torsion angles with linear time complexity, while ViS-MP enables efficient interaction between vector and scalar representations to fully utilize geometric information. ViSNet achieves state-of-the-art performance on multiple benchmarks, including the PCQM4Mv2 track in the OGB-LCS@NeurIPS2022 competition. It performs high-fidelity molecular dynamics simulations with reduced computational costs and high accuracy. ViSNet also demonstrates good interpretability and portability to other methods. It is effective for real-world applications, such as simulating the 166-atom mini-protein Chignolin, and shows improved performance compared to other models. ViSNet's design allows it to handle large molecular systems efficiently, making it suitable for applications in chemistry and biology. The model's performance is validated through extensive simulations and comparisons with other methods, demonstrating its effectiveness in molecular modeling and simulation.ViSNet is a geometric deep learning model designed to enhance molecular modeling by efficiently extracting geometric features and modeling molecular structures with low computational costs. It outperforms state-of-the-art methods on MD benchmarks like MD17, revised MD17, and MD22, and achieves excellent performance on QM9 and Molecule3D datasets. ViSNet efficiently explores conformational space and provides interpretable geometric-to-structural mappings. It uses an equivariant graph neural network (EGNN) approach, incorporating runtime geometry calculation (RGC) and a vector-scalar interactive message passing (ViS-MP) mechanism. RGC implicitly extracts geometric features like angles and dihedral torsion angles with linear time complexity, while ViS-MP enables efficient interaction between vector and scalar representations to fully utilize geometric information. ViSNet achieves state-of-the-art performance on multiple benchmarks, including the PCQM4Mv2 track in the OGB-LCS@NeurIPS2022 competition. It performs high-fidelity molecular dynamics simulations with reduced computational costs and high accuracy. ViSNet also demonstrates good interpretability and portability to other methods. It is effective for real-world applications, such as simulating the 166-atom mini-protein Chignolin, and shows improved performance compared to other models. ViSNet's design allows it to handle large molecular systems efficiently, making it suitable for applications in chemistry and biology. The model's performance is validated through extensive simulations and comparisons with other methods, demonstrating its effectiveness in molecular modeling and simulation.