We try to find the ICs of the system by accumulating data over a window, then update the initial condition of the window using an EAKF update with the data in a smoothing fashion. This is performed in place of the first backward-forward sweep. We can adjust the number of passes using n_initial_da_iterations where we repeatedly solve over the window and perform an update with the same data.
We are also using this PR to implement what appears to be necessary - regularization in the form of localization and inflation to prevent ensemble collapse.
TODO:
[ ] implement batching
[x] reduce computations to not require with full state matrices at every observation time in window
[x] create a new assimilation file which does only the iterated forward passes. (no backward-forward)
We try to find the ICs of the system by accumulating data over a window, then update the initial condition of the window using an EAKF update with the data in a smoothing fashion. This is performed in place of the first backward-forward sweep. We can adjust the number of passes using
n_initial_da_iterations
where we repeatedly solve over the window and perform an update with the same data.We are also using this PR to implement what appears to be necessary - regularization in the form of localization and inflation to prevent ensemble collapse.
TODO: