Lithology interpretation from well-log sequences using a depth-aware attention deep learning model

Aiming at lithology identification and stratigraphic interpretation for the L -well logging series in the N block of the Daqing Oilfield, this paper addresses common challenges such as ambiguous thin-bed boundaries, strong inter-well variability, and weak generalization of rule-based methods. This paper proposes an end-to-end lithology classification framework that fuses a 1D ResNet-18 backbone with a depth-aware attention Bi-LSTM. The proposed method takes windowed log sequences as input. The 1D ResNet-18 extracts multi-scale “morphological fingerprints” from multi-curve responses (e.g., spikes, step changes, attenuation patterns, and phase relationships), while channel recalibration adaptively emphasizes curve sensitivities varywith depth intervals. A Bi-LSTM then models vertical stratigraphic continuity using bidirectional context, and a depth-aware attention module focuses on key evidence around lithologic interfaces to produce continuous lithology tracks. To enforce geological plausibility, a depth-smoothing regularization is incorporated into the loss to suppress unrealistic high-frequency label oscillations. Experimental results demonstrate that the proposed method outperforms all baselines, achieving a Macro-F1 of 0.881 and an improved boundary detection performance with a boundary F1 of 0.842. These findings indicate that the proposed framework effectively balances local separability and vertical continuity, delivering interpretable lithology profiles that support well-to-well correlation, sand-body tracking, and fine-scale geological modeling.