Abstract
Nuclear magnetic resonance (NMR) data provides a comprehensive picture of the petrophysical description of a reservoir through effective characterization of fluid-rock properties. However, estimating the correct capillary pressure curves from NMR T(2) data in particular has been challenging with varied fluid saturation as it requires hydrocarbon correction. The earlier methods, either do not incorporate the hydrocarbon correction or exhibit limitations in their implementation, negatively impacting reservoir characterization. Therefore, in this work, a new methodology has been presented that estimates the P(c) in the reservoir at hydraulic flow units (HFUs) by using features of NMR T(2) and cumulative desaturation rate ∑(dS(nw)/dT(2)) through a newly developed workflow. Which incorporates the NMR hydrocarbon correction and encompasses the ensemble-committee machine model (ECMM) that has been purpose-formulated with the Bayesian optimized best-performing algorithms of machine, ensemble, and deep learning through a systematic approach. Results show that the ECMM workflow gives a much better mean squared error (MSE) than individual intelligent models while predicting P(c). ECMM has also been utilized to analyze the capillary pressure variability at HFUs which reveals that higher variance in capillary pressure values among HFUs cause model to underperform in terms of MSE and vice versa. The new methodology introduces a robust and cost-effective machine-learning incorporated workflow to estimate continuous capillary pressure for reservoirs having varied lithologies for effective characterization.