Spatial–Stratigraphic Information and Dynamic Range Attention Assist Well-Logging Lithological Interpretation

Time-series models, particularly CNN-BiLSTM architectures, have shown advances in the lithological interpretation of well-logging data. However, CNN and attention mechanisms face challenges in training efficiency and predicting precision. To improve this deficiency, a dynamic and lightweight attention mechanism and a strategy that combines geological/spatial information have been proposed. This study introduces two novel enhancements: the Spatial and Stratigraphic Information Processing (SSP, shorten as Spatial and Strat) method and the Dynamic Range Attention (DRA) mechanism. SSP integrates geological context by encoding depositional sequences as time series. DRA is a lightweight attention module that adaptively adjusts local attention range based on global context. Experiments on a collected dataset from eastern Sichuan Basin (13 wells, 14,587 labeled samples) demonstrate that the proposed DRA-BiLSTM model with SSP achieves excellent performance, achieving accuracies of 0.99 on the training set, 0.97 on the validation set, and 0.92 on the test set, with low error rates of 0.08 for Top-5 and 0.02 for Top-1. Ablation studies confirm the critical roles of SSP in capturing geological patterns and DRA in balancing computational efficiency by paying more attention to vertical sedimentary process. These innovations significantly advance automated lithological interpretation, offering a robust framework for geophysical applications.