Hi all. I’m Vaishak, coming from software and ML rather than geophysics, so please read this with that in mind.
Over the last few months I reproduced a published potential-field inversion and computed the posterior uncertainty on it: the closed-form linear-Gaussian case, on the Utah FORGE 3D gravity data (DOE GDR submission 1144, CC-BY 4.0). Code, figures and validation numbers are open: GitHub - vaishak-v-nair/Gurutva: Gravity, with error bars. Which parts of a subsurface model your survey actually supports - and a verdict that can refuse to answer. · GitHub
The part I would most like this forum to attack is not the posterior. It is the gates around it. Three results that surprised me.
Tuning the regularization weight until the misfit target is hit implies rock density varies by 0.021 g/cc. Real rock varies by roughly ten times that. Treated as a prior and tested against the data, that choice fails a licensing check at the 100th percentile: the observation is not a plausible draw from it. A flat independent-per-cell prior at a physical 0.25 g/cc fails too (99th), because uncorrelated draws are static, neighbours cancel, and the long-wavelength signal a real basin makes never appears.
On a Sleipner-class CO2 monitoring design, going from a 3 uGal to a 1.1 uGal gravimeter buys 6 percent. The interval is set by station geometry, not instrument noise. I expected the opposite and swept it to check.
Three self-consistency checks can all pass on a model that cannot reproduce its own data. On a satellite-gravity Moho model, licensing, calibration and stability all passed, and then a fourth check (does the recovered model actually reproduce the observations it was fitted to) failed. I withdrew the headline and published the failure. There is a fifth gate I have not solved, written up unresolved in the repo: coverage at a fixed realistic truth, versus declaring a prior class wide enough to contain the object you are searching for.
Because of that third point the tool is allowed to refuse. It returns one of NOT CLAIMED, PROVISIONAL, or CLAIMABLE, and on real data it returns PROVISIONAL rather than CLAIMABLE, because real data has no known right answer to check against.
My question. For those of you who invert potential-field data professionally: is a per-cell posterior standard deviation, plus an explicit “below this depth nothing is constrained by your data” line, actually useful to you? Or is the quantity that would change a decision something else entirely, such as a bound on an integrated mass, or a probability that a body exceeds some size, or something I have not thought of?
I would rather be told the whole framing is wrong now than after I build more on top of it.