Analysis & Interpretation
Feb 5, 2008
Don’t fall into the worst modeling trap! : "IF IT FITS
THE DATA, IT MUST BE RIGHT." (NOT SO!)
You can say: IF IT DOESN’T FIT IT’S WRONG (OR
AT LEAST INCOMPLETE).
Another trap with models: "If the model doesn’t fit the
data, keep adding variables." (NOT SO!)
THE NUMBER OF VARIABLES SHOULD BE
CONSIDERABLY LESS THAN THE DEGREES OF
FREEDOM OF YOUR DATA.
Decide what parameters you will allow to vary, and find
the best model you can. Don’t try to fit every data point
- just get the "most reasonable" fit.
• The gravity signal, corrected to FAA, contains information
from ALL mass in the earth - but only that relatively close
to you will yield a significant anomaly.
- How can you tell whether a signal is generated near
or far away?
• We adjust our gravity survey to emphasize anomalies
generated in the depth range we are concerned with.
• Remove "regionals" before working on shallower
Consider the following profile:
This profile contains at least three possible
anomalous sources. The SMOOTHED SLOPE of the
anomaly is likely the result of a deep structure. Our
profile is not long enough to resolve this anomaly, so
we can remove it by taking out a slope for the whole
We can do this in a modeling program by entering a deep
sloping layer that we can later ignore.
The anomaly still contains a bowl-shaped component that
likely is the result of either a broader or deeper
anomalous body. Since we can see much of this anomaly
(about one wavelength) we can attempt to model it.
The depth of this anomalous interface can be adjusted
along with the density contrast across the boundary to
provide reasonable geology - such as a valley buried
with sediment. Once that is done, the remaining
anomaly likely comes from smaller shallower sources:
Before adding the "bowl" to the model
After adding the "bowl" to the model
We are now in a situation where we can model the
SHORTEST wavelength (shallowest) anomalies without
interference from the longer wavelength, possibly
Is it worth using precise methods, such as
statistically accurate curve fitting, to model our
Probably not - since our models are necessarily non-
unique and arbitrary, why worry about precision? It
might tempt you to believe that your model is correct,
not just plausible.
What types of data can
constrain gravity models?