1460 The Leading Edge December 2009 SPECIAL SECTION: Reservoir modeling constrained by seismic Reservoir modeling constrained by seismic Revisiting the use of seismic attributes as soft data for subseismic facies prediction: Proportions versus probabilities G eostatistical modeling originated within the mining industry to estimate average minable ore grade from large support volumes given samples measured on small volume support.Inpetroleumgeostatistics,the goal is more equivocal due to several different scales of support of input data, which are often incongruent with the desired prediction scale. More specifically, the goal is to utilize indirect measurements (e.g., seismic data) from a scale larger than the prediction scale for fine-scale spatial distributions of facies and petrophysical properties grounded by undersampled point data (e.g., well- log data). (Note, volume support is a geostatistical term that describes the size or resolution of the sample or measurement.) For this modeling purpose, seis- mic attributes serve as soft data, i.e., imprecise information, which indi- rectly inform the three-dimensional spatial distribution of properties, while sparsely sampled well- log data better inform the vertical distribution of properties. Techniques for bringing these two disparate types of data into one model have been extensively addressed in the everyday application of petroleum geostatistics. However, in light of these advances, the original strength of geostatistical model- ing in addressing issues of scale or volume support has be- come secondary. he common practice for reconciling these disparate scales is to upscale log data and resample the seismic attri- butes to the model grid. he model grid is defined by depth- converted surfaces interpreted from seismic reflectivity pro- files and internal stratigraphic layers of a specified thickness between the surfaces. he layer thicknesses in the model grid are generally finer than the vertical seismic resolution, whereas the x-y cell size is usually the same as the binning of the seismic traces (Figure 1). For facies modeling, the sim- plest (not necessarily the best) upscaling of log facies uses a most of selection, also known as majority rule. Due to the indirect relationship between facies and seismic attributes, seismic attribute data are converted to a probability defining the presence or absence of a facies. Many different methods are used to generate these probabilities, from simple well- to-seismic calibration, to more detailed approaches such as LISA STRIGHT, ANNE BERNHARDT, ALEXANDRE BOUCHER , and T APAN MUKERJI, Stanford University RICHARD DERKSEN, Rohoel-Aufsuchungs AG principal component analysis (PCA), statistical rock physics, and neural network classification. he fundamental omission with these approaches is the lack of an explicit model of scale. hat is, the vertical resam- pling of both the well logs and the seismic attributes presup- poses a different volume support than is actually prescribed by the data set itself. he distribution and the spatial struc- ture of the data are inherently linked to the scale at which they were measured. Furthermore, a set of seismic attributes, of low resolution and band-limited frequency do not inform simply the presence or absence of a category; rather they in- form the combination of a specific pattern of subresolution categories that generate the resulting seismic response. his paper addresses the issues of modeling subseismic- scale facies with seismic attributes while explicitly incorpo- rating the information content of each datum at the correct support at which it informs the model scale. First, the scales of interest are introduced. hen a data-driven, multi-attribute, multiscale (MA-MS) well-to-seismic calibration is shown along with the application of the calibration to unsampled regions. his method addresses the issue of undersampling by using statistical rock properties and forward modeling. In an effort to test our methodology, the general workflow is demonstrated on a subsurface data set and the results linked to geostatistical and pattern-based facies modeling. SPECIAL SECTION: Reservoir modeling constrained by seismic Figure 1. Conceptual depiction of the input scales of data (seismic data, V, and well-log data, v) that contribute information for geostatistical modeling to property prediction at the model scale (m). Colocated log and seismic traces provide calibration information for using lower resolution, exhaustively sampled seismic attributes as soft information for facies proportion prediction. Downloaded 11 Dec 2009 to 171.64.170.81. Redistribution subject to SEG license or copyright; see Terms of Use at http://segdl.org/