Fig. 5: The comparison of model performance (y-axis), training time consumption (x-axis), and training memory consumption (volume) among ViSNet (red) and other algorithms (grey) including PaiNN, ET, MACE, GemNet-OC, Allegro, and NequIP on Chignolin. | Nature Communications

Fig. 5: The comparison of model performance (y-axis), training time consumption (x-axis), and training memory consumption (volume) among ViSNet (red) and other algorithms (grey) including PaiNN, ET, MACE, GemNet-OC, Allegro, and NequIP on Chignolin.

From: Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing

Fig. 5: The comparison of model performance (y-axis), training time consumption (x-axis), and training memory consumption (volume) among ViSNet (red) and other algorithms (grey) including PaiNN, ET, MACE, GemNet-OC, Allegro, and NequIP on Chignolin.

PaiNN and ET are faster and smaller as ViSNet further incorporates dihedral calculation. ViSNet outperforms GemNet-OC due to its Runtime Geometry Calculation, reducing the explicit extraction of dihedral complexity from \({{{{{{{\mathcal{O}}}}}}}}({{{{{{{{\mathcal{N}}}}}}}}}^{3})\) to \({{{{{{{\mathcal{O}}}}}}}}({{{{{{{\mathcal{N}}}}}}}})\). Additionally, ViSNet is also faster and smaller than MACE, Allegro, and NequIP for streamlining the CG-product. ViSNet achieves the best performance for its elaborate design, i.e., runtime geometric calculation and vector–scalar interactive message passing. Source data are provided as a Source Data file.

Back to article page