Fig. 4: Comparing predictions at 200ns for different values of the dynamical constraint 〈N〉.
From: Path sampling of recurrent neural networks by incorporating known physics
In this plot, we show the free energy profiles calculated from (a) the 100ns trajectory in the training set, (b) both the actual 200ns trajectory and direct prediction from LSTM, (c) the reference 200ns trajectory and ps-LSTM prediction with constraint of nearest-neighbor (NN) transitions 〈N〉 = 0.38, and (d) the reference 200ns trajectory and prediction with constraint 〈N〉 = 0.42. The table in (e) lists the kinetic constraint 〈N〉 calculated from corresponding trajectories. The averaged transition time τR→L and τL→R in picoseconds were calculated by counting the numbers of transitions in each trajectory. For reference MD, the error bars were calculated by averaging over transition time in a single 100ns or 200ns trajectory, while for the predictions from LSTM and ps-LSTM, the error bars were averaged over 10 independent predictions with the transition time for each predicted trajectory calculated in the same way as MD trajectories. The free energy profiles and the first NN values 〈N〉 in (b)–(d) are averaged over 10 independent training processes. The corresponding error bars are calculated as standard errors and filled with transparent colors.
