Figure 4.

(a,b) Performance comparison for additive BO (aBO), ordinary (oBO), and random search using the Li-tavorite dataset (averaged from 1000 trials). Horizontal axes for both denote the sequential number of DFT-$E_b$ evaluations $t$. The vertical axis in a) shows the residual gap at step $t$ between optimal $f\left({\mathit{x}}_{\mathbf{\ast }}\right)$ and the prior best solution ($ma{x}_{{\mathit{x}}_{1:t}}f\left({\mathit{x}}_t\right)$). The vertical axis in (b) shows the average probability ratio of discovering the optimal Li-tavorite compound ${\mathit{x}}_{\mathbf{\ast }}$. Note that both additive BO (aBO) and ordinary BO (oBO) search methods here used the tuned hyperparameters from half of the Li dataset excluded for search performance comparison (gray area).