Seismic Lithofacies Classification From Well Logs Using Statistical Rock Physics

In this paper we identify and classify populations of well log data into clusters, referred to as seismic lithofascies, representing sedimentary units with characteristic Seismic properties. We show that detailed analysis of local geology and physical rock properties can improve the understanding of variability in well log data and guide the selection of optimal well log parameters for facies discrimination and classification. We use data from a North Sea turbidity field, and define six different facies groups (I-VI) based on clay content, grain size, and bedding configuration. The facies are primarily determined from well logs (gamma ray, density, and sonic logs), but sub-facies of thick-bedded sandstones (Facies 11) are defined by certain textural parameters (clay location, cementation, etc.), which are determined from core and thin-section analyses. Rock physics modeling is used to guide the recognition of characteristic clusters of data. Having established a statistically representative training data base from a type-well, we perform multivariate classification of data from other wells in the area. We use different multivariate statistical methods and a neural network for the classification, and compare the success rates of the different methods. We find that the Mahalanobis discriminant analysis (MLDA), the probability density function (PDF) classification, and the neural network (NN) classification all have a "SUCCCSS rate" of about 80% when we use sonic and gamma ray logs together. The neural network does slightly better than MLDA, which again does slightly better than PDF. However, NN requires Much more computational effort than do MLDA or PDF. The advantage of PDF over MLDA is that it will easily reveal types of lithofacies other than those in the training data and/or detect erroneous log measurements. In general, this study shows that a relatively simple statistical technique as MLDA is effective for classification of well log data into distinct lithofacies with characteristic physical rock properties.