High-Fidelity saturation prediction using physics-informed attention neural network

Abstract

Accurate estimation of water saturation is critical for reservoir characterization and optimal production strategies in ultra-deep tight sandstone gas reservoirs. However, traditional empirical models often fail to provide reliable predictions due to the inherent heterogeneity and nonlinear interactions within these reservoirs. To address these challenges, this paper proposes the Saturation Neural Network (SatuNN), a novel deep learning framework that combines CNNs, window attention, axial attention and BiLSTM to capture multi-scale spatiotemporal features from logging data. Specifically, the CNNs and window attention layers capture local spatial features, while the BiLSTM and axial attention layers address global temporal dependencies. The proposed model leverages a physics-informed neural network strategy to embed petrophysical constraints directly into the training process, ensuring physically consistent and geologically meaningful predictions. Comprehensive evaluations using core and logging data from ultradeep tight sandstone reservoirs in the Tuha Basin demonstrate that SatuNN achieves superior predictive accuracy and significantly lower errors (R² = 0.92, MAE = 3.27 %) compared to both the optimal ablation baseline model (Without WA: R² = 0.86, MAE = 4.51 %) and traditional petrophysical model (Archie: R² = 0.84, MAE = 4.83 %). Moreover, successful field applications in two blind-test wells further validate the robustness and practical applicability of the proposed model. The presented SatuNN framework provides an accurate approach for saturation prediction in ultradeep tight sandstone gas reservoirs, effectively offering potential to improve reservoir evaluation and field development strategies.