Shortening an inversion

This case is typical of M1QN3. When a very small epsg is requested, the optimization may exit in mode 4: convergence is not reached but the maximum number of simulations is exhausted. By examining the gradient norm ratio in the logfile, you can identify a satisfactory iteration to stop at. Relaunch the job to stop at that iteration and generate the standard NetCDF outputs. To do this:

1. If directory fwd_-002 (which contains the final results) has already been created, remove it.

2. Modify the YAML to reload the available computations:

   • in obsoperator, set autorestart to True

   • in simulator, set reload_from_previous to True. If simulator is not specified in the YAML (the usual case when using the default cost function), add it as shown in the example below.

mode:
  minimizer:
    df1: 0.5
    epsg: 0.1
    maxiter: 10
    nsim: 25
    plugin:
      name: M1QN3
      version: std
    simulator:
      plugin:
        name: gausscost
        version: std
      reload_from_previous: true
  plugin:
    name: 4dvar
    version: std
  save_out_netcdf: true
obsoperator:
  plugin:
    name: standard
    version: std
    type: obsoperator
  autorestart: True

3. Set the convergence criterion (epsg for M1QN3) to stop at the targeted iteration. The value must be greater than the target gradient norm ratio but smaller than all previous ones.

4. If needed, rename the logfile so the details of existing simulations are not overwritten by the new run.

5. Launch the job.