Use posterior ensemble mean for new prior about flownet

[anders-kiaer]

The posterior parameters from an AHM run is automatically stored, and is already being used in prediction runs:

flownet/src/flownet/prediction/_run_pred.py

Lines 53 to 54 in 6447c6d

	with open(args.ahm_folder / "parameters.pickled", "rb") as fh:
	parameters = pickle.load(fh)

The same file could probably be used when solving this task.

A possible workflow could be:

• Grab the previously saved posterior parameter values (from the stored file mentioned above, already in use by pred).
• Fit to some prior distribution parameters.
• Modify* the setup here, where the prior parameters are currently initialized:
  

  flownet/src/flownet/ahm/_run_ahm.py

  Lines 358 to 376 in 2107fc5

  	parameters = [
  	PorvPoroTrans(porv_poro_trans_dist_values, ti2ci, network),
  	RelativePermeability(
  	relperm_dist_values, ti2ci, df_satnum, fast_pyscal=fast_pyscal
  	),
  	Equilibration(
  	equil_dist_values,
  	network,
  	ti2ci,
  	df_eqlnum,
  	equil_config.datum_depth,
  	config.flownet.pvt.rsvd,
  	),
  	RockCompressibility(
  	config.model_parameters.rock_compressibility.reference_pressure,
  	config.model_parameters.rock_compressibility.min,
  	config.model_parameters.rock_compressibility.max,
  	),
  	]

The modification would include then not only supporting creating dataframes with prior distributions (with distribution parameters from the config), but also from the stored posterior parameters. The dataframes already have support for defining prior distributions e.g. per tube, so the change necessary should not be huge 🎉

[olwijn]

All prior parameter distributions are currently defined by min and max (or one of these and a mean) and a boolean indicating if the distribution is loguniform or uniform. These prior distribution parameters are save in the file parameters.pickle. Actual values are sampled by ERT and updated values are written to a parquet file. In prediction mode, these values are read by the ERT job CREATE_FLOWNET_MODEL (_render_realization.py). Proposed solution: the parquet file for the final iteration of the previous history match should be read, the mean should be calculated, and new distribution parameters should overwrite the values in the pickle file.

[olwijn]

A new optional entry can be added to the config file under model_parameters (ahm_case) that specifies the path to a previous HM experiment:

The file parameters_iteration-latest.parquet.gzip should be located in ahm_case.
The ensemble-mean parameter values are computed and used to find a new extreme value for a uniform or loguniform distribution through optimization (the maximum value from the prior distribution is kept fixed if the mean has moved towards the minimum and vice versa).