Conventional reservoirs benefit from a long scientific history that correlates successful plays to seismic measurements through depositional, tectonic, and digenetic models. Unconventional reservoirs are less well understood, however benefit from significantly denser well control. Thus, allowing us to establish statistical rather than model-based correlations between seismic data, geology, and successful completion strategies. One of the more commonly encountered correlation techniques is based on computer assisted pattern recognition. The pattern recognition techniques have found their niche in a plethora of applications ranging from flagging suspicious credit card purchase patterns to rewarding repeating online buying patterns. Classification of a given seismic response as having a " good" or " bad" pattern requires a " distance metric". Distance metric " learning" uses past experiences (well performance) as training data to develop a distance metric. Alternative distance metrics have demonstrated significant value in the identification and classification of repeated or anomalous behaviors in public health, security, and marketing. In this paper we examine the value of three of these alternative distance metrics of 3D seismic attributes to the identification of sweet spots in a Barnett Shale play.
Similarity of waveforms or attribute patterns along a horizon to those about good and bad wells is a well-established practice in seismic data interpretation. In such studies, typically, the interpreter compares a vector of samples (e.g. Johnson, 2000) or attributes (e.g. Michelena et al., 1998) extracted from productive or non-productive wells to every trace along the horizon. Poupon et al. (1999) correlated wells to seismic waveforms, where the supervision was not only the actual seismic about the well but also a suite of synthetic seismic traces generated through petrophysical modeling and fluid substitution. The performance of these algorithms can depend delicately on the manner in which distances are measured. Distance metrics are a vital component in many applications ranging from supervised learning and clustering to product recommendations and document browsing. Since, designing such metrics by hand is difficult, we look at the problem of learning a metric from exemplars. In particular, we consider relative and qualitative exemplars of the form " P is closer to Q than P is to R". In essence, a distance metric is a mathematical operator that conveys how similar two (possibly vector-valued) members of a set compared with a single, scalar value, based on a notion of similarity. The most common and most easily understood waveform similarity metric is based on Euclidian distance.