Figure 3:

Effect of potential roughness on the performance when using a loss function based on the infinitesimal lag time limit. Results shown are obtained with (22) and (23) with $\omega =0$ (top row), and $\omega =10$ (bottom row). (A) The potentials. (B) Reference committors obtained from the finite-difference scheme in [38, 60]. (C) Neural network prediction of the committors. (D) Differences between the references and the predictions. (E,F) Same as columns C and D, except for a lag time of 100 steps. Note the different scales on the difference maps in columns D and F.