Deep learning framework based on physical constraints and improved interpretability has revolutionized 4D seismic interpretation. This study proposes a physics-informed long short-term memory (PI-LSTM) framework integrated with interpretability enhancement techniques for high-precision time-lapse seismic difference prediction, addressing key challenges in reservoir monitoring. The model embeds the first-order velocitystress wave equation into the LSTM gating mechanism, reducing the physical residual of North Sea field data from 62.3 kPa to 15.2 kPaa 75.6% decrement. An interpretability enhancement module combines Shapley additive explanation value dynamic weighting with physical attention templates, reducing the seasonal fluctuation of feature importance by 38% (measured as S). Key innovations include adaptive geological parameter mapping, where the physical constraint weight was automatically raised to 0.89 0.04 when porosity exceeded 15%. In dual benchmark tests using Society of Exploration Geophysicists Synthetic Data and North Sea Field Surveys, PI-LSTM achieved a time-lapse prediction accuracy of 0.712.1 ms, equivalent to a hydrocarbon interface localization error of 35 MPa), the model reduced the wave equation residual to 18.6 kPa, 67.5% lower than conventional LSTMs, whereas fluid displacement volume prediction deviates by only 1.8% from well data. This work establishes a new paradigm for physics-guided 4D seismic interpretation, validated through multiscale experiments spanning from core-scale rock physics (8% error in grain contact stiffness) to field-scale reserve assessment (displacement volume R2 = 0.94).
Zhao et al. (Mon,) studied this question.