Figure5

Figure 5. Trajectories of the traversed structure during the search of different CSP algorithms. (A-C) show the trajectory for SrTiO$$ _3 $$; and (D-F) show the trajectory for MnAlCuPd. The trajectories were drawn by calculating the distance metrics for the valid structures during the search and mapping them into $$ 2 $$D space using t-distributed stochastic neighbor embedding (t-SNE). Two consecutive points were connected if the latter structure had a lower energy than the former one; (G) and (H) show the t-SNE for all three algorithms in the same figure for SrTiO$$ _3 $$ and MnAlCuPd, respectively. The initial and optimal structures for all algorithms are marked with various colors and shapes. The points in ParetoCSP's trajectory are more spread out and have more diverse search directions than the other algorithms.