fields

[rowanc1]

Sometime we pass around fields? Do we need to?

[lheagy]

In the current implementation we need to... for the data misfit it is important, otherwise they will be recomputed: https://github.com/simpeg/simpeg/blob/master/SimPEG/DataMisfit.py#L147

[lheagy]

We could think about attaching the fields to the simulation object in future implementations

[rowanc1]

I think that this pulling out is showing where some of the holes are. Fields should be closer to where they are used. The objective functions really shouldn't care that there are dependencies to an operation. These should be captured elsewhere (like the simulation).

I think that we would also have difficulty in the current implementation for a multiphysics inversion? if fields get passed to both objects? Not sure if @thast got around this or it was fine.

[thast]

I do not remember having to change anything and have not notice significant problem.
Here is an example on how I work with 2 data misfit:

Load Data

survey_grav = PF.Gravity.readUBCgravObs(
    "./GRAV_Synthetic_data.obs"
)
survey_mag = PF.Magnetics.readMagneticsObservations(
    "./MAG_Synthetic_data.obs"
)

Wires and mapping

actv = Utils.surface2ind_topo(mesh, topo, gridLoc='N')
actvMap = Maps.InjectActiveCells(mesh, actv, ndv)
wires = Maps.Wires(('den', actvMap.nP), ('sus', actvMap.nP))
gravmap = actvMap * wires.den
magmap = actvMap * wires.sus

Grav problem

prob_grav = PF.Gravity.GravityIntegral(mesh, rhoMap=wires.den, actInd=actv)
# Pair the survey and problem
survey_grav.pair(prob_grav)

Mag problem

prob_mag = PF.Magnetics.MagneticIntegral(mesh, chiMap=wires.sus, actInd=actv)
# Pair the survey and problem
survey_mag.pair(prob_mag)

Data Misfit

dmis_grav = DataMisfit.l2_DataMisfit(survey_grav)
wd_grav = np.ones_like(survey_grav.dobs) * 0.01
dmis_grav.W = 1 / wd_grav
 
dmis_mag = DataMisfit.l2_DataMisfit(survey_mag)
wd_mag = np.ones_like(survey_mag.dobs) * 1.
dmis_mag.W = 1 / wd_mag
 
dmis = dmis_grav + dmis_mag

[rowanc1]

Thanks for the code snippets!

I think this works because they are both linear problems and you don't have to account for the fields in your Jvec calls.

https://github.com/simpeg/simpeg/blob/master/SimPEG/PF/Gravity.py#L124

So although it is probably passing the fields object around, it is never used - and I am not sure which one it would be (mag or grav problem).

[thast]

Thanks for pointing this out @rowanc1 ! True I have not try joint inversion for nonlinear problem yet. I will keep you updated once I've tried it.