Effects of expanding paraDomain in ERT inversion

[benbpete]

Hi All,

I've been playing around with ERT modeling and came across an effect I don't understand. Below are two models (full paraDomain extents, and then a smaller scale figure) of the same dipole-dipole data set, the only difference is when creating the PLC/Mesh. For the small scale figures, I adjusted the color scales to be the same. Finally attached is the apparent resistivity psuedosection.

For the first model I used the default paraBoundary =2, and for the second I used paraBoundary = 24. The result in the final model is pretty large. Both models converge very quickly with chi^2 of 0.18 and 0.1.

The inversion is simply:
mod = ert.invert(data, mesh=mesh, zWeight=1, lam = 100, verbose=True, robustData = False)

I can attach my code/data if it makes any difference. Any thoughts?

Thanks!
Ben

[halbmy]

All the models are equivalently fitting the data and show a blurred two-layer model. Outside of the area of resolution the model is arbitrary and depends on resolution.

However, with a 100m surface layout you will not be able to resolve anything beyond 30m depth, so the models are all too deep (and some too wide). Seems like the alpha-shading heuristics makes only sense for models of reasonable size.

[benbpete]

Yeah, what I am doing is just exploring how changes in the paraDomain size affect the final model. It was good to see how using the sensitivity as a guide for model limits really changes depending on the size of the paraDomain and that it isn't reliable with overly large modeling domains.

Just so I am thinking about this correctly, even though the models are equivalently fitting, the reason for the differences is likely a combination of 1) larger paraDomain which results a larger distance to the model boundaries and 2) change in triangular mesh size near the left/right boundaries of the paraDomain and 3) likely a large number of equivalent solutions because the resistivity contrast in the "true model" is relatively small and gradational. The sensitivity is also spread out over a large area, which I am guessing results in a larger "domain" of equivalent solutions. Anything else that might be affecting it?