Publications

In the Abadi field, permeability prediction from well logs has been one of the important subsurface evaluation tasks to evaluate field-wide dynamic reservoir performance. Toward this end, K-nearest neighborhood algorithm (K-NN), one of the most popular non-parametric regression techniques, was employed in the previous study (Zushi et al., 2011). The technique achieved good fitting quality for core permeability, however significant discrepancy from well test derived permeability has been recognized as one of the issues to be solved. Based on the thorough review of available data and methodology, the followings were raised as possible main causes of the discrepancy. (1) Insufficient core coverage in the high permeability zone and algorithmic limitation of the K-NN. If high permeability data is missing in the data base, the prediction by the K-NN algorithm, a local averaging method, provides the value within the data range and results in underestimation. (2) Inappropriate training of the non-parametric model. Each depositional facies has its own characteristic porosity-permeability relationship in this field, however the relationship may not be appropriately captured if the algorithm is applied to all the data at once. To avoid the possible causes, we adopted projection pursuit regression technique (PPR) (Friedman et al., 1981) and constructed a permeability prediction model by separately training for the individual rock types defined by petrophysical characteristics associated with the geological framework. In this paper, we explain the geological framework defined in the Abadi field and the application procedure of PPR algorithm for permeability prediction, and demonstrate the significant improvement in terms of both fitting quality for the core data and consistency with well test evaluation. The overall procedure can be easily applied to other fields.