Adaptive weight strategy for frequency-decomposed seismic attribute fusion in predicting of complex sand body distributions

High-precision sand prediction is fundamental to improving the efficiency of oil and gas exploration and development. To address the limitations of traditional fixed-weight fusion strategies, particularly under conditions of significant lateral variation in sand body distribution, this study proposes a dynamic weighting–deep neural network (DW-DNN) for adaptive frequency-decomposed attribute fusion. The approach integrates physical constraints with deep learning and introduces two innovations: (ⅰ) a priori weight matrices derived from the amplitude–frequency and tuning thickness relationship (amplitude variation with frequency, AVF) are embedded into the attention mechanism to adaptively allocate multiband seismic attributes, emphasizing high-frequency features for thin sands and low-frequency features for thick sands; and (ⅱ) a deep neural network with a composite loss function combining mean squared error (MSE) and AVF-based constraints is designed to jointly optimize weight allocation and prediction accuracy. The method was applied to the Xi 233 area of the Qingcheng Oilfield in the Ordos Basin and compared with conventional approaches. DW-DNN achieved high accuracy and generalizability, with an R² of 0.92 in the 30% blind-well test, 24.3% higher than conventional methods. In addition, 91% of well-point errors were within 0–3 m, while prediction accuracies for thin (≤3 m) and thick (>3 m) sands reached 88% and 91%, respectively. The model also maintained stable performance under low well-control conditions (training–test ratio 5:5). Predicted sand distributions exhibited improved continuity and geologically plausible geometries, clearly delineating channels, lobes, and estuary bars. The results demonstrate that DW-DNN enhances frequency-decomposed attribute fusion through adaptive weight allocation, providing a robust tool for predicting sand body distributions in complex reservoirs.