Spatiotemporal seismic attribute prediction via SA-GNN and physics-constrained LSTM

Abstract The accuracy of seismic attribute prediction directly affects the interpretation of underground structures and the identification of oil and gas reservoirs. In view of the difficulty of traditional methods in modeling non-Euclidean spatial dependencies and physical dynamics, this paper proposes a joint prediction method that integrates self-attention graph neural network (SA-GNN) and physical constraint LSTM (PG-LSTM). SA-GNN uses a multi-head self-attention mechanism to dynamically allocate node weights, significantly improving the spatial feature expression ability of complex geological structures. The velocity prediction error in the fault area is reduced by 46.2% compared with the traditional LSTM (MAE from 13.82 m/s to 7.43 m/s); PG-LSTM embeds a one-dimensional wave equation regularization term to ensure that the time series evolution conforms to the physical law, reducing the phase prediction error by 22.4% (MAE from 3.87 rad to 3.01 rad). Experiments show that the mean absolute error (MAE) of this method in velocity, amplitude and phase prediction on the SEG SEAM Phase I dataset is 7.43 m/s, 5.49 (dimensionless) and 3.01 rad respectively, the coefficient of determination (R²) is increased to 0.943, 0.893 and 0.871, and the structural similarity (SSIM) is 0.89. In addition, the model remains robust under small sample (25% training data) and high noise (-10 dB) conditions, and the error volatility is reduced by 35% compared with the baseline. This study provides a new modeling framework for seismic attribute prediction that combines high accuracy and physical consistency, which can be extended to application scenarios such as intelligent inversion and real-time monitoring.