Reserve Estimation Based on Post-Stack Seismic Inversion via Multilayer Perceptron, Random Forest, and Extra Tree Regression Algorithms

This study presents an enhanced reserve estimation approach that mitigates the limitations of conventional inversion methods, thereby facilitating more informed decision-making during exploration. The training process used acoustic impedance (AI) logs as the target variable and depth, two-way travel time, and seismic attributes at the well log scale as inputs. The models employed multilayer perceptrons, random forests, and extra tree regressions to predict the AI, which was then compared to the BLIMP (band-limited impedance inversion) data. The methodology emphasizes selecting suitable cutoffs or thresholds of porosity and permeability for optimal performance. Using cutoff values that match porosity and permeability extremes reduces error and improves reservoir parameter estimation for net pay. The hydrocarbon pore volume is subsequently computed, followed by the estimation of reserves and economic value. Compared with the available cited ground truth data (1.8310e+11 STB and 3.9549e+12 SAR), the ETR results in more accurate reserves and values (1.7174e+11 STB and 3.7096e+12 SAR). These results are also supported by the ETR's empirical value (a=0.1584) being closest to the value of the ground truth (a=0.1585). We conclude that, compared with the ground truth, the ML models exhibit varying performance, with the accurate trend of ETR > RFR > MLPR. In contrast, the values of the BLIMP are smaller than those of the ETR, provided that the threshold choices meet the optimum criteria. The novelty of this study lies in the introduction of a new way of obtaining reserve estimation based on the ML approach of poststack seismic inversion with good accuracy. Seismic attributes are treated as multiple seismic traces, which serve as inputs to ML models to provide predictive information for reserve estimation.