Digital core permeability is a crucial factor in rock engineering, reservoir simulation, and underground applications, serving as the foundation for evaluating fluid flow underground. However, predicting digital core permeability using artificial intelligence presents challenges such as high model calculation complexity, strong resource dependence, and low efficiency. To overcome these challenges, this paper presents a novel hybrid enhanced rock topology-guided graph neural network (RTG-GNN) to accurately characterize the pore structure of digital cores and predict permeability. This model takes advantage of the physical attributes of pores and throats, utilizing them as features for nodes and edges. It improves information exchange through a gating mechanism and incorporates an attention mechanism for weighted aggregation of neighboring data, culminating in accurate permeability predictions. Furthermore, a scheme for constructing a comprehensive permeability distribution dataset (CPDD) is proposed for digital cores. This scheme effectively mitigates challenges like inadequate training data, subjective sampling bias, and data drift. Multiple experimental results demonstrate that RTG-GNN surpasses most methods in various performance indicators. Additionally, the RTG-GNN model boasts a lightweight design, occupying only 17% to 44% of the size of mainstream models and requiring just 15% to 38% of GPU resources for training. Remarkably, while maintaining high accuracy and low error rates, the computational load is reduced by a factor ranging from 19.06 to 636.41, the training speed is enhanced by 12.12 to 86.47 times, and the inference speed is increased by 6.76 to 16.14 times. The experiment further confirms that both single core and mixed rock datasets, acquired through the CPDD scheme, exhibit excellent generalization and reliability.
Read full abstract