AUTOMATIC FLOW-CALIBRATED PERMEABILITY LOG ESTIMATION THROUGH DUAL RANDOM FOREST ALGORITHM: HOW PREDICTIVE ANALYTICS ENHANCES THE CHARACTERIZATION OF HETEROGENEOUS CARBONATE RESERVOIRS

In carbonate reservoirs, the estimation of a reliable permeability log is a long-standing problem mainly because of the inherent multi-scale heterogeneities. The conventional approach relies on core-calibrated algorithms applied to open-hole (OH) logs. In general, this static log-based prediction uses to underestimate the actual dynamic performance of the wells and an ad-hoc integration with production logging tool (PLT) and well test (WT) analyses represents a required step to correct the initial estimation. However, it is critical, and at once challenging, to define the relation between dynamic-based corrections and OH characterization outcomes. An elegant solution is here proposed that makes use of predictive analytics applied on special core analyses (SCAL), nuclear magnetic resonance (NMR) log modeling, and multi-rate PLT/WT interpretations. The methodology is presented for a complex oil-bearing carbonate reservoir and it starts with an advanced NMR characterization performed downhole for more than 100 wells, and after a rigorous calibration with SCAL. The main outputs are a robust porosity partition (in terms of micropore, mesopore and macropore contributions), and a physics-based permeability formula. Although the match with core data demonstrates the reliability of the applied NMR rock characterization, log permeability underestimates the actual dynamic performances obtained from WT, as expected. At the same time, multi-rate PLT data from more than 150 wells are used to compute an apparent permeability value for each perforated interval, automatically consistent with the associated WT interpretation. Finally, both static and dynamic characterization outputs are used as inputs for a dual random forest (RF) template. In detail, the first RF algorithm learns through experience how NMR porosity partition and core-calibrated permeability are related to PLT/WT apparent permeability values, after considering the proper change of scale. Next, the second RF is utilized to estimate the uncertainty associated to the previous step, still in a completely data-driven way. Hence, the so-defined dual model provides a continuous automatic flow-calibrated permeability log, together with its confidence interval, directly from static NMR responses. The presented methodology allows dynamic data to be incorporate efficiently into a static workflow by means of a pure data-driven analytics approach. The latter is able to shed light on the statistical relationships hidden in the available datasets, thus leading to a more accurate permeability estimation. It is also shown how this provides fundamental information for perforation strategy optimization and reservoir modeling purposes in such carbonate rocks.