Gas Saturated Sandstone Reservoirs Characterization through Probabilistic Seismic Inversion Using Bayesian Neural Networks

Abstract

Shallow gas-saturated sandstone reservoirs are prevalent throughout Indonesia, presenting highly lucrative economic prospects that are often hindered by complex geological heterogeneities. Characterization of these reservoirs typically relies on deterministic seismic inversion methods. However, the inherent non-uniqueness of such approaches yields singular solutions devoid of uncertainty quantification. This study proposes a robust probabilistic seismic inversion methodology utilizing a Convolutional Bayesian Neural Network (BNN). By incorporating Monte Carlo (MC) dropout, which forces the stochastic deactivation of hidden nodes during training, the algorithm successfully generates a comprehensive ensemble of predictive realizations. The proposed framework is evaluated using 2D pre-stack synthetic angle gathers (0 to 45°) across two scenarios: a baseline layer-cake model and a structurally complex faulted model incorporating random noise to simulate field-acquired data. Using seismic amplitudes and low-frequency models to simultaneously predict porosity, clay volume, and water saturation, the BNN demonstrates exceptional predictive accuracy on the baseline model, yielding an average Root Mean Square Error (RMSE) of 0.03 and a coefficient of determination (R²) of 0.9. When applied to the complex faulted model, the predictive fidelity naturally degraded (average R² of 0.7 to 0.8; RMSE of 0.09 to 0.15), consistent with the inherent challenges of noise-contaminated data. Furthermore, a comparative analysis against a conventional model-based deterministic inversion reveals that while the deterministic approach better delineates fault architectures via horizon constraints, the BNN framework yields superior quantitative fidelity. Notably, the BNN successfully overcomes the systematic underestimation of porosity inherent in the deterministic acoustic impedance-to-porosity conversion process. The stochastic inversion explicitly produces 2D petrophysical cross-sections detailing the P10, P90, and mean distributions extracted from 50 realizations. Given their high fidelity, the mean distributions are directly applicable for reservoir characterization. Crucially, the probabilistic P10 and P90 boundary percentiles facilitate a comprehensive assessment of in-place hydrocarbon volumes. This study are currently limited to synthetic validation, and subsequent validation with real-field data is necessary before the approach can be practically applied.