Enhanced fracture network permeability prediction using attention mechanism and Kolmogorov–Arnold Networks with SHAP interpretability analysis

Accurate prediction of the equivalent permeability of fractured reservoirs is important for reservoir characterisation and subsequent reservoir-scale simulation. This study proposes a hybrid framework (Attention-enhanced Kolmogorov–Arnold Network, AKAN) for predicting the equivalent permeability of fractured reservoirs. The framework integrates discrete fracture network modelling (DFNWORKS) with graph-theoretic connectivity analysis to construct the dataset. It leverages attention mechanisms to dynamically weight critical geological features, while the Kolmogorov–Arnold Network (KAN) serves as a non-linear prediction core, enhanced by attention parameters to capture hierarchical feature interdependencies. Benchmarking against baseline models (KAN, CNN–KAN, and CNN–LSTM), AKAN achieved superior convergence efficiency (MSE = 0.13), with MAE, MAPE, MSE, and RMSE reduced by 51.38–57.45%, 40.80–41.37%, 65.10–65.82%, and 58.87–59.01%, respectively, in predicting the equivalent permeability. SHAP analysis reveals that the interaction feature of fracture intensity and radius distribution coefficient contributes the highest importance, with a cumulative SHAP value magnitude accounting for approximately 68.2% of the total. Furthermore, using transfer learning, the model was effectively deployed to a fractured reservoir site in Gansu, where it achieved a permeability prediction accuracy of 99% and a residual of 0.41 mD. This interpretable framework effectively combines predictive accuracy with feature-level interpretability, offering a reliable tool for permeability assessment in fractured reservoirs.