Fig. 3
From: Multiobjective genetic training and uncertainty quantification of reactive force fields
Evolution of the Pareto optimal front solution (red) in a 10th generation, b 50th generation and c 260th generation, compared with the local Pareto optimal solution in 1st generation (blue). The solution converges to the global Pareto optimal solution in 260 generations as shown in c. Each of the Mo–O, H–S and Mo–S bond errors is calculated as \(\mathop {\sum}\nolimits_{i = 1}^{N{\mathrm{frames}}} {(BondCount_{{\mathrm{RMD}}}\left[ i \right] - BondCount_{{\mathrm{QMD}}}[i])^2}\), where BondCountRMD[i] and BondCountQMD[i] are the number of bonds in the ith time frame estimated using RMD and QMD simulations, respectively (N frames is the total number of time frames). Convergence is quantified in d, e and f, where the sum of errors (divided by 10,000) is plotted for all the points in each generation. The dashed blue lines show the average error in the first generation, whereas the dashed red lines show the average error in 10th, 50th and 260th generations, respectively. Lowering of the dashed red lines with successive generations shows convergence of the proposed scheme
