Building the carbonate pore-type classifer for well logging via the blended training dataset

The exploration of carbonate rocks has outstanding economic benefts, as well as facing the extreme challenge of reservoir characterization. This article has proposed a data-based description scheme generalizing carbonate pore-type characteristics from both laboratory measurements and theoretical predictions to the well logging dataset. Firstly, in the feature space of elastic properties, we employed the supervised machine learning (ML) algorithm to convert this pore-type classifcation process from a typical nonlinear inversion to sample label allocation problem. Secondly, to alleviate the inherent scale gaps between data sources, virtual samples were randomly mixed into the laboratory measured dataset. Through inheriting or mimicking statistical elastic features of limited core samples, the new built training dataset could improve the overall sample richness and thus help the ML algorithms making better identifcation decisions. On the one hand, this scheme was verifed by 74 carbonate samples. In the feature space of high dimensions, the blended dataset trained radial basis function support vector machine accurately separated diferent carbonate pore systems. Moreover, using logging curves of a carbonate gas feld, we verifed the generalization capability of this scheme over unbalanced data scales. Searching skills were used to optimize model and classifer setups according logging curves of a specifc interval. Finally, with the help of the vertical label distributions, logging elastic response modes and historical pore evolution footprints were further studied.