INTEGRATION OF PETROPHYSICAL ANALYSIS WITH NEURAL NETWORKS.
THE TRAINING AND ESTIMATION PHASE.
Pattern recognition in the formation evaluation process is as important
as geostatistics. It helps to add a supplementary model to the
multiple realizations of a reservoir analysis.
Neural networks in petrophysics are used in supervised learning mode.
In some static geological models applications, neural networks
are used as an aid to facies prediction.
Petrophysicists use such applications to help where mineralogical
models fail to fully explain the petrophysical properties or to try
and forecast data in unexplored areas.
In order to get the full functionalities of neural networks applications
for petrophysical interpretation the analyst should be aware of the
parameter interdependence for the data involved in the estimation
process.
The input and output training data set should be geostatistically
representative for the specific formations and a sensitivity analysis
is a priority for the selection of the input curves.
Sensitivity analysis or the dependence of a petrophysical property from
other properties is a problem that should be solved before the neural
network training phase.
Normally running multiple realizations, the NN application recognizes
the right weight to be given to a specific curve in order to predict
a petrophysical property.
It is convenient to use the curves that have a major direct influence on
the simulated log curve.
For the sonic curve simulation priority is to be given to the density curve.
If we choose as input curves GR, ILD and RHOB we have a different
DSCP curve as with GR, ILD.
Esteem. Neural Networks applications for Log Analysis. ( Courtesy of DigitalFormation )
Estimation of DTSC from RHOB and GR, input and output curves selection
Courtesy of DigitalFormation
Output results example from RHOB and GR
Courtesy of DigitalFormation
Evaluation of DTSC from RHOB, GR and RILD, input parameters
Courtesy of Digitalformation
Evaluation of DTSC sonic curve from GR, RILD and RHOB