mlcolvar.utils.timelagged.create_timelagged_dataset

class mlcolvar.utils.timelagged.create_timelagged_dataset(X: Tensor, t: Tensor = None, lag_time: float = 1, reweight_mode: str = None, logweights: Tensor = None, tprime: Tensor = None, interval: list = None, progress_bar: bool = False, walker: Tensor = None)

Bases:

Create a DictDataset of time-lagged configurations.

In case of biased simulations the reweight can be performed in two different ways (reweight_mode):

1. rescale_time : the search for time-lagged pairs is performed in the accelerated time (dt’ = dt*exp(logweights)), as described in [1] .

2. weights_t : the weight of each pair of configurations (t,t+lag_time) depends only on time t (logweights(t)), as done in [2] , [3] .

If reweighting is None and tprime is given the rescale_time mode is used. If instead only the logweights are specified the user needs to choose the reweighting mode.

References

[1]

Y. I. Yang and M. Parrinello, “Refining collective coordinates and improving free energy representation in variational enhanced sampling,” JCTC 14, 2889-2894 (2018).

[2]

J. McCarty and M. Parrinello, “A variational conformational dynamics approach to the selection of collective variables in meta- dynamics,” JCP 147, 204109 (2017).

[3]

H. Wu, et al. “Variational Koopman models: Slow collective variables and molecular kinetics from short off-equilibrium simulations.” JCP 146.15 (2017).

Parameters:

• X (array-like) – input descriptors

• t (array-like, optional) – time series, by default np.arange(len(X))

• reweight_mode (str, optional) – how to do the reweighting, see documentation, by default none

• lag_time (float, optional) – lag between configurations, by default = 10

• logweights (array-like,optional) – logweight of each configuration (typically beta*bias)

• tprime (array-like,optional) – rescaled time estimated from the simulation. If not given and reweighting_mode`=`rescale_time then tprime_evaluation(t,logweights) is used

• interval (list or np.array or tuple, optional) – Range for slicing the returned dataset. Useful to work with batches of same sizes. Recall that with different lag_times one obtains different datasets, with different lengths

• progress_bar (bool) – Display progress bar with tqdm

• walker (array-like, optional) – Identifier of the trajectory (walker) to which each configuration belongs. This can only be used when reweight_mode is set to weights_t.

Returns:

dataset – Dataset with keys ‘data’, ‘data_lag’ (data at time t and t+lag), ‘weights’, ‘weights_lag’ (weights at time t and t+lag).

Return type:

DictDataset

__init__(**kwargs)