Fig. 2
From: De novo exploration and self-guided learning of potential-energy surfaces
“Learning” the crystal structure of α-rhombohedral boron. Top: Error of iteratively generated GAP-RSS models, for the energy of the optimized ground-state structure of α-B12, referenced to DFT. Three independent runs are compared: random selection of points (gray), our two-step selection procedure with CUR on SOAP vectors (green), and the same two-step procedure but with CUR on SOAP similarity kernels (purple). Bottom: Evolution of the B12 icosahedron as the defining structural fragment. For three points of the N = 100 cycles, having completed 400 (“A”), 500 (“B”), and 2500 (“C”) DFT evaluations in total, the respective lowest-energy structure (at the DFT level) from this iteration is shown, as visualized by using VESTA.64 Bonds between atoms are drawn using a cutoff of 1.9 Å; note that there are further connections between the B12 icosahedra with slightly larger B· · ·B distances
