Lithofacies Clustering Using Principal Component Analysis and Neural Network: Applications to Wireline Logs
- PDF / 3,094,330 Bytes
- 19 Pages / 439.37 x 666.142 pts Page_size
- 98 Downloads / 214 Views
Lithofacies Clustering Using Principal Component Analysis and Neural Network: Applications to Wireline Logs Y. Zee Ma
Received: 19 July 2010 / Accepted: 14 March 2011 / Published online: 3 May 2011 © International Association for Mathematical Geosciences 2011
Abstract Both statistical methods and artificial neural network (ANN) have been used for lithology or facies clustering. ANN, in particular, has increasingly gained popularity for clustering of categorical variables as well as for predictions of continuous variables. In this article, we discuss several counter examples that show deficiencies of these techniques when used for automatic lithofacies clustering. Our examples show that the lithofacies clustered by ANN alone or ANN in combination with principal component analysis (PCA), as commonly used, are highly inconsistent with the benchmark charts based on laboratory results. We propose several techniques to overcome these problems and improve the clustering of lithofacies, including (1) classification of lithofacies using the minor or intermediate principal component(s), (2) rotation of a principal component before using ANN for clustering, (3) cascading two or more PCAs and ANNs for clustering lithofacies or electrofacies, and (4) classifying lithofacies with demarcated stratigraphic reference classes. Keywords Rotation of principal component · Minor component · Cascaded clustering · Artificial neural network · Geologic interpretation · Reservoir characterization · Stratigraphic reference class 1 Introduction The importance of geologic facies or lithology in reservoir modeling rests upon their governance of petrophysical properties and subsurface fluid flow. The term lithofacies is generally used in this article, which may connote geologic facies or lithology depending on the context. Lithofacies is an intermediate scale of heterogeneity and has a duality of being an attribute and reference class in reservoir characterization Y.Z. Ma () Schlumberger Ltd., 6501 S. Fiddler’s Green Circle, Suite 400, Greenwood Village, CO 80111, USA e-mail: [email protected]
402
Math Geosci (2011) 43: 401–419
and modeling (Ma et al. 2009). Yet unlike many petrophysical properties, lithofacies are not measured as wireline logs, and their interpretation from core data is generally limited. In order to populate lithofacies in a reservoir model, they are predicted at the wells using wireline logs that describe rock properties. As a matter of fact, since the introduction of wireline logs more than eight decades ago, they have been the main data sources for rock formation evaluation, including lithofacies interpretation. However, as each wireline log has different characteristics in its depth of investigation and sensitivity to the rock properties (Tilke et al. 2006), an abundance of wireline logs often makes interpretation of the lithofacies difficult. Several methods have been used for predicting lithology or facies by using wireline logs. Early methods applied cutoffs on wireline logs to derive them. For example, it is a common pr
Data Loading...