Efficient Prediction of Fracture Density with Limited Labeled Samples: A Deep Active Learning Framework
Аннотация
Identifying fractures in complex reservoirs from conventional well logs remains challenging due to limited labeled samples, heterogeneous fracture distributions, and the lack of uncertainty quantification in deterministic models. Accordingly, a deep active learning framework integrating Bayesian inference (AL-BiLSTM) is proposed for fracture density (FVDC) prediction. First, a three-level feature selection strategy comprising mutual information (MI), principal component analysis (PCA), and dynamic adaptive Lasso [least absolute shrinkage and selection operator] (DA-Lasso) is designed to eliminate the multicollinearity of conventional log responses and extract core features. Second, a Bayesian bidirectional long short-term memory (BiLSTM) network with a Monte Carlo (MC) dropout mechanism extracts the long-range spatial dependencies of well logs while quantifying the epistemic uncertainty of model predictions. Finally, maximum epistemic uncertainty is utilized as the acquisition function for well-level active sampling. This guides the model to iteratively query the most informative candidate wells, thereby reducing reliance on expensive FMI logs. The framework is verified using a volcanic reservoir dataset from the Junggar Basin under a leave-one-well-out cross-validation protocol. Using only two to three training wells, the framework achieves prediction performance comparable to or better than that of the passive-learning baseline trained on all four non-test wells. When active querying exhausts the three-well candidate pool, yielding a four-well training set, the AL-BiLSTM attains an root mean square error [RMSE] of 0.0897 on test well X1.