mchoblet / ensemblefilters Goto Github PK

For the code in ensrf_direct.py check if eliminating the matrix inversions results in a speed up (inverting matrices instead of solving linear systems is inefficient)
Functions to test:

• np.solve
• scipy.linalg.cho_solve (using that the covariance matrices are positive definite -> Cholesky decomposition)

Additionally add an option for a distant dependent covariance localization which is directly added in the function as a schurproduct (as in the DASH package)

Some little example on how to use it with the ensrf_direct_loc function.
(Just needs some model_data with coordinates (xarray-format) + proxy locations.)

Probably also really fast. Covariance localization should be strait forward as for the direct solver, and as we use large ensembles
the statistical part of perturbing the observations should not be a problem.