INTEGRATING SEISMIC INVERSION PRODUCTS INTO MULTI-ATTRIBUTE ANALYSIS AND UNSUPERVISED MACHINE LEARNING FOR RESERVOIR CHARACTERIZATION, POHOKURA, TARANAKI BASIN, NEW ZEALAND

Abstract

Recent advances in machine learning have enhanced the capability of multi-attribute analysis to resolve relationships within high-dimensional data, enabling more sophisticated reservoir characterization. However, the ability of multi-attribute analysis to produce geologically meaningful results remains strongly controlled by the diversity, relevance, and spatial extent of the input data. Seismic amplitude data is inherently band-limited and does not adequately capture low-frequency components, which limits its ability to represent subsurface variability. Inversion products derived from low-frequency models provide a complementary source of information to address this limitation.This study investigates the effects of incorporating pre-stack simultaneous AVO inversion products into multi-attribute analysis. To evaluate the impact of inversion products, a comparative framework based on two input datasets is established. The first scenario utilizes attributes derived solely from seismic amplitude data, whereas the second incorporates P- and S-impedance volumes into the attribute set. The Pohokura Field in the Taranaki Basin is used as a case study, where multi-attribute analysis is performed for both scenarios using Principal Component Analysis (PCA) and K-means clustering. The results indicate that the inclusion of inversion products leads to a shift in data representation from a structure primarily governed by seismic amplitude response to one that is more consistent with subsurface geology and controlled by elastic properties. This transition improves the delineation of reservoir and seal units and enhances the definition of reservoir boundaries, lithological trends, heterogeneity zones, structural discontinuities, and facies distributions. Sand-prone intervals that are not clearly resolved in the attribute-only scenario become detectable when inversion products are included. Furthermore, the inclusion of inversion products results in reduced noise and improved vertical and lateral resolution, which enhances interpretability and enables a more reliable characterization of the complex Mangahewa reservoir. The achievements gained by incorporating inversion products into multi-attribute analysis highlight the importance of enriching the frequency content of the input data. In this regard, multi-attribute analysis should be regarded not simply as a data integration approach, but as a design-oriented process in which the selection and organization of the input dataset fundamentally control the quality of the results.