Geological interpretation - 3-D Seismic Interpretation - page 112

Output

Input

Fig. 4.5 Schematic neural network (after Gurney, 1997).

pick a top and base surface of the body by the usual manual methods, and then
display the seismic between these surfaces as a semi-transparent 3-D body.
(6) An entirely different approach is to use a classification algorithm to map areas of
similar seismic character. One approach to this is to use neural net software. A
general introduction to the principles of neural networks may be found in Gurney
( 1997 ) , for example. A simple example of such a network is shown in fig. 4.5 .
Input data are fed to nodes in the first layer of the network. Each arrowed path has
associated with it a weight, and the input values are multiplied by the weight cor-
responding to the particular path that they travel. On arrival at a node, the weighted
inputs are summed; if the sum exceeds a threshold value, the node sends out a
high signal value (conventionally '1') to the next layer of nodes, or if the threshold
value is not reached the node sends out a zero signal value. The same weighting
procedure is carried out along the paths to the second layer, and the nodes in that
layer sum the inputs and outputa1ora0depending on whether the summed input
exceeds the threshold or not. This behaviour of the nodes mimics that of nerve cells
(neurons) in biological brains. The weights on the interconnecting paths determine
how the system behaves. They can be determined by a learning process, in which
input data are presented for which the correct output is known. For example, if we
wanted to predict whether a certain layer is sand or shale from seismic data, then
the input could be a set of seismic attribute values (trace amplitudes, loop widths,
etc.) at a well location where it was known whether the layer was sand or shale. If
we had a number of wells, some with sand and some with shale, each with its own
seismic attribute values from an adjacent trace, it would be possible to adjust the
neural net weights iteratively so that the output is a sand/shale flag when the seismic
attribute values are supplied as input. An extension of this idea is to predict values
of reservoir properties such as porosity. One way of doing this is to classify trace
data according to their similarity to synthetics produced from wells with known