Well‐Log Lithology Classification in Continental Clastic Reservoirs Using Structural Feature Tokenization and Attention‐Gated Heterogeneous Model Fusion

ABSTRACT Accurate lithology identification from wireline well logs underpins reservoir evaluation in continental rift basins, where sandstone, siltstone, and mudstone produce overlapping petrophysical responses—a condition termed spectral equivalence—that defeats conventional threshold‐based interpretation. Systematic baseline offsets between wells within the same field further constrain the generalizability of existing supervised models. This study proposes LGB‐Transformer, a hybrid deep learning framework that addresses spectral equivalence, class imbalance, and probability miscalibration as a coupled system and explores its generalization performance under cross‐well distributional shift through multi‐well joint training. The framework integrates a LightGBM branch, repurposed as a structural feature tokenizer that encodes hierarchical petrophysical decision boundaries through leaf–node indices, with a Transformer encoder branch that models global cross‐curve dependencies through self‐attention. A context‐conditioned attention gate generates per‐sample weighting coefficients to dynamically balance the two branches. Training employs Focal Loss to redirect gradient mass toward hard‐to‐classify boundary samples, and a post hoc Platt scaling step is included as a standard calibration procedure, though quantitative evaluation shows its calibration improvement is limited on this low‐dimensional class‐imbalanced well‐log dataset. Evaluated on 38,225 depth‐indexed records from four wells in a continental rift basin clastic reservoir, LGB‐Transformer achieves an overall accuracy of 89.98%, an area under the receiver operating characteristic curve of 0.9719, a Cohen's kappa coefficient of 0.81, and a macro‐averaged F1 score of 87%—consistent improvements over seven baselines spanning gradient boosting, kernel, and deep learning methods. Vertical prediction profiles confirm superior boundary fidelity in thin‐layer alternating sequences relative to all comparison models. The results demonstrate that combining gradient boosting structural priors with Transformer global context modeling, together with a two‐stage reliability procedure, provides a practical and reproducible pathway to automated lithological interpretation in geologically complex, low‐dimensional well‐log environments. Leave‐one‐well‐out validation further reveals that zero‐shot cross‐well generalization remains a challenging task under limited well samples and severe intra‐well class imbalance. Therefore, large‐scale validation efforts will be actively pursued in the future, incorporating more wells, external basin datasets, and richer logging datasets to test and enhance the method's general applicability across various continental clastic reservoir types.