fig2

Figure 2. Replacing MLPs in different parts of the ML potential Allegro^[10] with KANs employing various basis functions. Z_i stands for the chemical species of atom i. $$ \overrightarrow{r_{i j}} $$ stands for the relative displacement vector from atom i to atom j. Substituting MLPs with KANs generally enhances prediction accuracy. Specifically, replacing MLPs in the output block of the Allegro model results in higher prediction accuracy and shorter training time than using MLP, and higher inference speed and higher computation resource efficiency than using KANs throughout the entire model. MLPs: Multi-layer perceptrons; KANs: Kolmogorov-Arnold Networks; ML: Machine learning.