InterPore2026

Physics-Informed LSTM Network for Water Saturation Prediction in Heterogeneous Tight Sandstone Reservoirs: Integrating Petrophysical Constraints with Sequential Data

19 May 2026, 15:05

1h 30m

Poster Presentation (MS15) Machine Learning in Porous Media Poster

Speaker

Sha Li

Description

Accurate prediction of water saturation (Sw) is paramount for evaluating reserves and productivity in tight sandstone gas reservoirs. However, the strong heterogeneity, complex pore-throat structures, and high clay content of these reservoirs pose significant challenges for traditional petrophysical models (Archie’s equations) and pure data-driven machine learning (ML) methods. While ML models capture non-linear features, they often lack physical consistency and rely heavily on large-scale labeled datasets, which are scarce in deep, tight formations. In this study, we propose a novel hybrid framework, the Physics-Informed Long Short-Term Memory (PI-LSTM) network, designed to predict Sw with both high accuracy and physical interpretability. This approach integrates the sequential learning capability of LSTM—to capture the depth-dependent geological trends—with the physical constraints of porous media theory. Specifically, we embed Archie’s Law and Darcy’s Law-based fluid distribution principles into the loss function of the neural network. This ensures that the model’s outputs adhere to fundamental petrophysical bounds and relative permeability relationships, even in intervals with sparse or noisy logging data. The proposed model was validated using well-log and core data from complex tight sandstone reservoirs. Compared to the traditional Archie model and standard Random Forest (RF) and LSTM algorithms, the PI-LSTM model demonstrated superior performance: (1)Enhanced Accuracy: The Root Mean Square Error (RMSE) of Sw prediction was reduced by approximately 18.5% compared to the Archie method and 9.2% compared to pure LSTM. (2)Physical Consistency: Unlike pure data-driven models that occasionally produced non-physical values (Sw > 1 or sudden fluctuations in homogeneous zones), the PI-LSTM maintained results within strictly plausible petrophysical ranges (0＜Sw＜1). (3)Data Robustness: In scenarios with a 50% reduction in training samples, the PI-LSTM maintained a high R2 (>0.88), while standard ML models showed significant performance degradation.Our findings suggest that incorporating physical laws into deep learning frameworks can significantly mitigate the "black-box" nature of AI in reservoir characterization. This PI-LSTM approach provides a robust and efficient tool for evaluating fluid distribution in heterogeneous porous media, offering a new perspective for the intelligent development of unconventional energy resources.