Machine learning-driven acoustic impedance inversion with globally optimized reservoir characterization for reserve estimation in carbonate reservoirs

• Predicted acoustic impedance (AI) was generated using multilayer perceptron regression (MLPR), random forest regression (RFR), and extra tree regression (ETR) models compared to AI log and band-limited impedance inversion (BLIMP). • Optimized porosity ( Φ ) and permeability ( K ) derived from the AI data were obtained using empirical exponential rock physics relationships and petrophysical power-law transforms. • Net pay determination based on the condition that the optimized Φ and K values exceed their thresholds. • Net pay masking was applied to identify the productive zone within the hydrocarbon pore volume before computing the original oil in place and performing reserve estimation. • The ETR provided the closest match to the ground truth of the AI log data and all downstream results until the reserve estimation. Reserve estimation in carbonate reservoirs remains challenging because of the nonlinear, power-law relationships among the acoustic impedance (AI), rock physics, and petrophysical properties that govern net pay identification. This study develops a machine learning (ML)-based seismic inversion workflow that integrates multilayer perceptron (MLPR), random forest (RFR), and extra tree regression (ETR) algorithms to predict AI from seismic attributes, depth, and two-way travel time. A limited-memory Broyden-Fletcher-Goldfarb-Shanno with Boundaries (LBFGS-B) optimization minimizes the square errors between reservoir extremes and the threshold values to obtain empirical constants under defined initial values and bounds. A ± 3 % perturbation to the initial values produced the global minimum errors. The solutions of these empirical constants generate optimized porosity and permeability profiles that determine the net pay when the computed reservoir properties exceed their thresholds. Further results show that ML-based inversion provides more accurate reserve estimates than conventional band-limited impedance inversion methods do. Due to hyperparameter choices among the tested models, the ETR achieved the closest match to the true empirical values for the AI-porosity transformation and reserve estimate (a ETR = 0.1584, R ETR = 28.08 million stock tank barrels (STB)) relative to the ground truth (a true = 0.1585, R true = 29.96 million STB), confirming its robustness for evaluating the Ilam carbonate reservoir. The proposed workflow demonstrates the potential of ML for enhanced reservoir characterization and hydrocarbon reserve estimation in complex carbonate systems.