Collocated cokriging and neural-network multi-attribute transform in the prediction of effective porosity: A comparative case study for the Second Wall Creek Sand of the Teapot Dome field, Wyoming, USA

• Collocated cokriging and neural network were compared for porosity prediction.
• Collocated cokriging overpredicted but failed to predict large porosities.
• Neural network underpredicted but better mimics log porosity.
• Neural network performed better than collocated cokriging.

Collocated cokriging (CCK) and neural-network multi-attribute transform (NN-MAT) are widely used in the prediction of reservoir properties because they can integrate sparsely-distributed, high-resolution well-log data and densely-sampled, low-resolution seismic data. CCK is a linear-weighted averaging method based on spatial covariance model. NN-MAT, based on a nonlinear relationship between seismic attributes and log values, treats data as spatially independent observations. In this study, we analyzed 3-D seismic and well-log data from the Second Wall Creek Sand of the Teapot Dome field, Wyoming, USA to investigate: (1) how CCK and NN-MAT perform in the prediction of porosity and (2) how the number of wells affects the results. Among a total of 64 wells, 25 wells were selected for CCK and NN-MAT and 39 wells were withheld for validation. We examined four cases: 25, 20, 15, and 10 wells. CCK overpredicted the porosity in the validation wells for all cases likely due to the strong influence of high values, but failed to predict very large porosities. Overprediction of CCK porosity becomes more pronounced with decreasing number of wells. NN-MAT largely underpredicted the porosity for all cases probably due to the band-limited nature of seismic data. The performance of CCK appears to be not affected significantly by the number of wells. Overall, NN-MAT performed better than CCK although its performance decreases continuously with decreasing number of wells.