I feel that the data used here has been stretched to the limits of its inherent usefulness. In my estimation, the following attributes (not an exhaustive list) describe ``ideal'' data for this problem:

- 1.
- Dense, wide-area seismic coverage.
- 2.
- 20 or more wells inside seismic coverage.
- 3.
- Five or more picked seismic horizons.
- 4.
- Access to stacking velocities.

I combine ``Conclusions'' and ``Future Work'' into one section, because SEP
is in
the process of acquiring a large group of North Sea data from Mobil that will
better
allow us to develop and test new ideas, thanks to a collaboration with the
**Stanford Center for Reservoir Forecasting** (SCRF). The data consists of
five densely sampled seismic horizons and over twenty wells with well log data.

Verification of the validity of the final model is
a difficult and ill-defined task. The method of *cross-validation* seems
most promising. The optimization problem of Equation (8) is
solved, but one of more of the well logs is disregarded. At the disregarded
well locations, compare the depth to the actual horizon from the well log and
that predicted by
. Since the first and second
``ideal data attributes'' should be satisfied to make this scheme viable, I do
not utilize it
in this paper, but I think the Mobil data is well-suited for the task.
One
can imagine a more complicated cross-validation scheme, in which many values
of are determined from Equation (8)
by removing different well logs. The result is a map of cross-validation
errors, which is then minimized in another least-squares scheme.

7/5/1998