Lithofacies Identification via Hierarchical Dilated Multi-Grained Scanning Model Integrating with Variational Feature Denoising based on Well Logging Data

Reliable lithofacies identification is pivotal for stratigraphic evaluation, reservoir characterization, and drilling assessment. Data-driven logging-based methods are widely used for lithofacies prediction, but conventional methods suffer from noise interference, reducing blind well identification accuracy. To address this, a hierarchical dilated multi-grained scanning model integrated with variational feature denoising (DMGSVD) is proposed. First, logging data are denoised via a variational autoencoder with a redesigned loss function. Then, a hierarchical classification model with dilated multi-grained scanning is constructed to fully exploit denoised lithology characteristics. Comparative experiments on onshore Daniudi and Hangjinqi datasets and offshore Norwegian North Sea datasets show DMGSVD’s average identification accuracy outperforms RF, LightGBM, LSTM, and TabNet by 1.15%-2.02%, 1.17%-2.02%, 8.51%-17.3%, and 3.53%-7.55% respectively. It effectively eliminates logging data noise and exhibits superior blind well identification performance under complex sedimentary conditions, demonstrating high practical value.