Rock permeability, as a significant parameter, plays a crucial role in research related to geological exploration, reservoir resource development, and the movement and distribution of subsurface fluids. Despite significant advancements in artificial intelligence technology for rock analysis in recent years, challenges remain in terms of prediction accuracy, computational complexity, and resource dependency. To address these issues, this paper proposes a superpixel-based graph neural network (SP-GNN) method to achieve rock permeability prediction. Specifically, the flow space of the rock medium is superpixelized to construct graph data. Based on the superpixel graph data, a graph neural network designed to learn multi-level cascading features is developed, while also implementing an algorithm to extract neighborhood spatial features that contain sequential relationships. A temporal pooling strategy is proposed to perform collaborative learning of the cascading features and neighborhood spatial features of graph data from a temporal perspective, in order to obtain hierarchical global features of dynamic nodes. Finally, the global features of the graph are input into the downstream tasks to achieve accurate prediction of permeability. Experimental results show that the SP-GNN significantly outperforms various existing benchmark schemes for permeability prediction across seven comprehensive performance metrics, demonstrating the ability to accurately and efficiently predict permeability for both mixed-type and single-type rocks; ablation experiments further validate the effectiveness of the proposed temporal pooling strategy.
扫码关注我们
求助内容:
应助结果提醒方式:
