Fig. 1: Procedure for path sampling LSTM.
From: Path sampling of recurrent neural networks by incorporating known physics
This schematic plot shows the workflow for constraining some static or dynamical variable s(Γi), given an unconstrained LSTM model. The workflow begins with generating numerous predicted trajectories from the constraint-free LSTM model. The corresponding variables that we seek to constrain can be calculated from the predicted trajectories and are denoted by s(Γ1), s(Γ2), s(Γ3) in the plot. We then perform path sampling and select a smaller subset of trajectories in a biased manner that conforms to the desired constraints, with a probability \(P(s({\Gamma }_{i}))\propto {e}^{-\Delta \lambda s({\Gamma }_{i})}\), where Δλ is solved by the Eq. (6). The subset is then used as a new dataset to train the LSTM model.
