ConvNeXt Block-Based Deep Learning Surrogate Model for Polymer-Flooding Reservoir Simulation

Surrogate modeling has emerged as a crucial strategy for time and cost savings in reservoir development. However, the majority of research on surrogate modeling has been concentrated on secondary reservoir development processes such as water or gas flooding, with limited focus on surrogate modeling for tertiary oil recovery techniques like polymer flooding. In this paper, we emphasize the utilization of deep modeling techniques and introduce a deep encoder-decoder neural network to establish a high-dimensional nonlinear dynamic mapping surrogate model for uncertainty parameters associated with polymer flooding. Specifically, this model maps uncertain grid permeability parameters to reservoir state variables such as pressure, saturation, and polymer concentration. The model employs an encoder network to extract high-dimensional coarse features of input parameters, followed by a decoder network to refine these features to obtain the final model outputs. Experimental results from applying this model to the dynamic prediction of a two-dimensional synthetic polymer flooding model show improved prediction accuracy compared to two existing surrogate models, while simultaneously reducing computational time for reservoir simulators, thus maintaining prediction precision.