PREDICTION OF POROSITY DISTRIBUTION USING MULTI-ATTRIBUTE PROBABILISTIC NEURAL NETWORK (PNN) AT VOLVE FIELD, NORWAY

Abstract

Predicting porosity distribution from seismic data in geologically complex areas, such as the Volve Field, is a significant challenge due to the ambiguous and non-linear relationship between rock properties and seismic attributes. This study directly deals with this issue by proposing and validating a multi-attribute Probabilistic Neural Network (PNN) as a method to generate a more accurate porosity model for the Middle Jurassic Hugin Formation. The methodology integrated data from two wells with a post-stack seismic volume. Key steps included performing a model-based inversion to create an acoustic impedance volume, using stepwise regression to identify the four optimal seismic attributes for porosity prediction, and training the PNN to learn the relationship between those attributes and the well log porosity. The cross-validation was employed throughout the process to prevent overfitting and ensure the model quality. The results quantitatively confirmed the improvement of the PNN approach. The network as a non-linear method achieved improvement as its training correlation of 0.965 with 9.9% average error, over the linear regression result of 0.942 correlation and 12.6% average error. The validation of PNN also gain slight improvement resulted a correlation of at 0.898 with an rage error of 16.3%, an improvement over the 0.881 correlation and 19.7% error from a linear regression. Following this validation, the trained network was applied to the entire dataset, resulting in a comprehensive porosity distribution volume. In conclusion, this study validates the multi-attribute PNN as a more reliable method than conventional linear techniques for the specific task of predicting porosity distribution, demonstrating the improvement of this approach to effectively model the distribution of the porosity in geophysical data.