Well-Logging Identification of Shale Lithology via FAtt-CNN: A Case Study From the Lianggaoshan Formation, Sichuan Basin

Why It Matters

Accurate shale lithology classification reduces exploration risk and boosts hydrocarbon recovery, offering a competitive edge for oil‑and‑gas operators.

Key Takeaways

• FAtt-CNN adds adaptive feature weighting to well‑log data.
• ND and PR envelopes identified as top discriminative log parameters.
• Model reaches 91.95% lithology identification accuracy in Sichuan Basin.
• Outperforms cross‑plot methods and prior deep‑learning benchmarks.
• Enhances detection of thin sand‑shale interbeds and transitional lithologies.

Pulse Analysis

Well‑log interpretation has long struggled with overlapping signatures in continental shale reservoirs, limiting the precision of reservoir models. Conventional cross‑plot methods rely on static thresholds, which often misclassify thin interbeds and transitional lithologies. Recent advances in machine learning, particularly deep‑learning architectures, promise dynamic feature extraction, yet many models lack mechanisms to prioritize the most informative log responses, leading to sub‑optimal generalization across diverse geological settings.

The Feature‑Attention Convolutional Neural Network (FAtt‑CNN) addresses this gap by embedding a parameter‑adaptive weighting module that amplifies lithology‑sensitive signals. In the Lianggaoshan Formation case study, researchers selected the neutron‑acoustic time‑difference envelope (ND) and the photoelectric‑resistivity ratio envelope (PR) as primary inputs, recognizing their high discriminative power. The model’s attention layer dynamically adjusts the influence of each feature, enabling it to capture subtle variations that differentiate thin sand‑shale interbeds from homogeneous shale. Validation results show a 91.95% identification accuracy, outperforming both traditional cross‑plots and competing deep‑learning models in recall and precision.

For the oil and gas industry, such a leap in lithology classification translates into more reliable reservoir characterization, optimized well placement, and reduced drilling uncertainty. By reliably isolating thin productive layers, operators can better target completion strategies and improve net present value calculations. The success of FAtt‑CNN also signals broader applicability of attention‑driven neural networks in other subsurface domains, from mineral exploration to carbon‑capture site assessment, positioning AI as a cornerstone of next‑generation geoscience workflows.