fig3

From: Robust global optimization of atomic structures via a learning loss-informed on-the-fly firefly algorithm

Figure 3. The active learning procedure implemented in NARA. The green-colored boxes in the middle column represent energy evaluations performed using the GNN, while the orange-colored boxes in the right column correspond to energy evaluations based on the target potential (e.g., DFT). NARA: Nature-inspired algorithm for robust atomic structure search; GNN: graph neural network; DFT: density-functional theory; SCF: self-consistent field calculations; LCB: lower-confidence-bound.