| Abstract | The study area is located in North Pattani Basin, Gulf of Thailand. Although the major
wells in the interested area did have logging and coring, not all depth intervals could be
cored and not all logging curves were measured. Some logs were even missing.
Assessment or prediction of petrophysical parameters, including porosity in uncored
intervals and at other wells locations may have an impo1tant role in the reserve
estimation. To help so lving this issue, Back-propagation Artificial Neural Network
(ANN) with early stopping method was employed in this study to help estimate porosity
from well logging data or seismic attributes.
Well logging data and core measurement of two wells (wells A and B) were studied.
Quicklook interpretation was conducted in order to identify reservoir, estimate shale
volume, porosity, permeability, and water saturation. The Back-propagation ANN with
early stopping method was studied and applied for porosity prediction by using Matlab
software. The data sets were separated into three subsets, i.e. training, validating, and
testing. The input data were from well logging data or seismic attribute data. The output
data used for training were from core porosity. The ANN model with least performance
error was applied for porosity prediction in uncored well and other wells locations.
The first group of ANN analyses used the well logging data as input layer with the
standard set consisting of GR, LLD, RHOB, NPHI, and DT. The desired output data were
core porosity measurements. The tangent hyperbolic function was employed as transfer
function for the hidden layer, while the linear function was used as transfer function for
the output layer. A total of 56 ANN analyses were performed to predict porosity for
uncored intervals. It was found that the ANN-based porosity of the best model matched
very well with core porosity, MSE was 10.09. Within this first group of ANN analyses, it
is wo1th mentioning that a number of analyses were done for the case the sonic log (DT)
was missing. To solve this problem of missing DT, the synthetic DT curves were
reconstructed using both methods of Log Response Equation (LRE) and ANN. As a good
finding of this study, the DT curves were well reconstructed, MSE was 80.41. Again,
based on the reconstructed DT, a complete input data layer including GR, LLD, RHOB,
NPHI, and DT, was obtained for the ANN analysis, MSE was 9.22.
The second group of ANN analyse was conducted with the major change in the input
layer, with the input data being se ismic attributes (dip, azimuth, instantaneous phase, and
relative acoustic impedance). These seismic attributes were extracted from seismic cube
by Petrel software. A total of 6 ANN analyses were run to study the seismic attribute
sensitivity. The matching of ANN-based porosity was not that good as those obtained in
the first group. The best results obtained were obtained for the case using all seismic
attributes as input data, MSE was 44.30. However, the result was encouraging in the
sense that if seismic attributes could be successfully used as the input data in an ANN
analysis one has good opportunity to make a satisfactory 3D petrophysical modelling (or
3D distribution of porosity in this case) for the concerned oil field. |