Spatiotemporal neural network with attention embeddings as surrogate for reservoir automatic history matching

Neural network surrogate modeling offers a viable alternative to the computationally intensive numerical simulations typically used in reservoir history matching. However, developing surrogates that maintain high accuracy while remaining adaptable to realistic reservoir conditions remains a significant challenge. Motivated by practical reservoir applications, this study proposes a task-oriented, module-enhanced neural network surrogate. By incorporating a spatiotemporal architecture with separable self-attention and the Cauchy activation function, the proposed model facilitates rapid, accurate, and end-to-end prediction of production dynamics in realistic reservoir systems. The developed neural network surrogate model is coupled with a geometric inflation factors-based Ensemble Smoother with Multiple Data Assimilation, forming an efficient surrogate-assisted framework for rapid reservoir automatic history matching. Validated on two field-scale reservoirs representing distinct displacement mechanisms (waterflooding and CO2 flooding), the proposed model achieves excellent predictive accuracy for production data, with loss values below 0.023 and coefficients of determination (R2) exceeding 0.947. Compared with conventional methods, the proposed framework achieves comparable inversion accuracy while significantly enhancing computational efficiency, demonstrating its efficacy in realistic reservoir conditions.