NEDAS.assim_tools.assimilators.serial module

class NEDAS.assim_tools.assimilators.serial.SerialAssimilator(c: Context)

Bases: Assimilator

Subclass for serial assimilation algorithms

assim_mode: str = 'serial'

init_partitions(c: Context) → list

Generate spatial partitioning of the domain

assign_obs(c: Context)

Assign the observation sequence to each partition par_id

Parameters:

c (Context) – the runtime context object

Returns:

Indices in the full obs_seq for the subset of obs that belongs to partition par_id

Return type:

dict[ObsRecordID, dict[PartitionID, np.ndarray]]

property loop_cost: CallCost

The three per-observation calls of the serial loop, timed (NEDAS/utils/call_cost.py; opt-in with NEDAS_CALL_COST, otherwise the functions are called untouched). Shared by every serial assimilator, so the native EAKF and a compiled one are timed at the same boundaries; what the loop spends outside them is the broadcast, the distance computations and bookkeeping, which are the same Python for every backend.

distribute_partitions(c: Context)

Distribute partitions across processors

assimilation_algorithm(c: Context)

Implementation of the serial assimilation algorithm.

Notes

serial assimilation goes through the list of observations one by one for each obs the near by state variables are updated one by one. so each update is a scalar problem, which is solved in 2 steps: obs_increment, update_ensemble

abstractmethod obs_increment(obs_prior, obs_prior_static, obs, obs_err) → ndarray

Compute observation-space analysis increments.

Parameters:

• obs_prior (np.ndarray) – Observation priors, 1-D float array of length nens

• obs_prior_static (np.ndarray) – Observation priors of the static members (covariance_def.nens_static)

• obs (float) – The real observation value

• obs_err (float) – Observation error std

Returns:

observation-space analysis increments

Return type:

ndarray

abstractmethod update_local_state(state_prior, state_static, obs_prior, obs_prior_static, obs_incr, ind, state_h_dist, state_v_dist, state_t_dist, hroi, vroi, troi, h_local_func, v_local_func, t_local_func, correlation_local_func, impact_on_variable) → None

Update the local state vector with the analysis increments.

Parameters:

• state_prior (np.ndarray) – Local state vector, point-major, shape (nloc, nfld, nens)

• state_static (np.ndarray) – Local state of the static members, shape (nloc, nfld, nens_static), not updated

• obs_prior (np.ndarray) – Observation priors, shape (nens,)

• obs_prior_static (np.ndarray) – Observation priors of the static members, shape (nens_static,)

• obs_incr (np.ndarray) – Analysis increments, shape (nens,)

• ind (np.ndarray) – The candidate points (indices into nloc) the distances are given for

• state_h_dist (np.ndarray) – Distances, shapes (nind,), (nfld, nind), (nfld,)

• state_v_dist (np.ndarray) – Distances, shapes (nind,), (nfld, nind), (nfld,)

• state_t_dist (np.ndarray) – Distances, shapes (nind,), (nfld, nind), (nfld,)

• impact_on_variable (np.ndarray) – Cross-variable localization factor per variable, shape (nfld,)

abstractmethod update_local_obs(obs_data, obs_data_static, used, obs_prior, obs_prior_static, obs_incr, ind, h_dist, v_dist, t_dist, hroi, vroi, troi, h_local_func, v_local_func, t_local_func, correlation_local_func, impact_on_variable) → None

Update the local observations with analysis increments.

Parameters:

• obs_data (np.ndarray) – obs prior ensemble, point-major, shape (nlobs, nens)

• obs_data_static (np.ndarray) – obs priors of the static members, shape (nlobs, nens_static), not updated

• used (np.ndarray) – boolean mask of already-assimilated obs

• ind (np.ndarray) – The candidate obs (indices into nlobs) the distances and impact_on_variable are given for

class NEDAS.assim_tools.assimilators.serial.BoxBins(grid, x, y, w)

Bases: object

Neighbor search as DART’s get_close: the points (x, y) are sorted into bins about w wide once, then candidates() returns the points in the bins that overlap grid.search_box, so distances are computed for those alone. Bins count from the grid corner; along a cyclic dim they split the period evenly and wrap around. Every point is a candidate when the grid gives no box.

candidates(ref_x, ref_y, r)

Indices of the points that may lie within distance r of (ref_x, ref_y).