The Petroleum Software Technology neural network prediction software, NNLAP, is used to estimate log data were none was collected.  This is often very desirable especially since bulk density and porosity logs are seldom run in all logging runs and since synthetic seismograms often depend as much on bulk density changes as on seismic velocity changes.  An example of a case where a bulk density log is extrapolated in a well based on the gamma ray log:

The fit of this technique is based in this case on four zones indicated by the vertical green lines in the figure above.  Shown in the figure in the bottom track are the gamma ray log used to predict the bahavior of the density log (GR), the predicted gamma ray (GRNN), the collected density log in blue in the middle track (RHOB) and the estimated density log (RHOBNN).  Also plotted in the top track are the sonic (white) and deep induction (green) logs.  The fit measured and estimated RHOB curves for the four data zones are shown below:

Note that errors within a small distance in depth for these four zones are small as indicated in the second figure.  In most cases, some tuning is necessary in selection of points used to 'train' the neural network.  Since each zone was chosen to be 11 depth points wide, these four zones are not just 4 simple values.  Selection of the width of zones and their placement allows an analyst to omit portions of a log which represents hole-diameter problems or washouts or other problems with the logs as invasion of the drilling mud into the formation. Such a problem is noted in the bulk density log between depths shallower than 3175.  This technique allows us to use the estimated log at shallower depths where no density log was collected.