Boundary-Guided U-Transformer for Automatic Stratigraphic Correlation from Multichannel Well Logs

Abstract Fine-scale stratigraphic correlation supports reservoir characterization, but manual interpretation is time-consuming and may vary among interpreters. This study proposes a Boundary-Guided U-Transformer (BG-UT) for automatic stratigraphic correlation from multichannel well logs. The method formulates correlation as a one-dimensional semantic segmentation task along the depth direction. A U-shaped convolutional encoder–decoder with skip connections is combined with a Transformer bottleneck incorporating relative positional bias, enabling the network to integrate local well-log response patterns with information from a broader depth interval. Auxiliary branches for stratigraphic top/base regression and boundary detection are jointly optimized with the classification task, providing complementary supervision for interval geometry and interface positions. During inference, sequence-constrained Viterbi decoding combines classification scores with transition preferences based on the expected vertical succession, encouraging stratigraphically coherent label sequences. The decoded labels are then converted into contiguous stratigraphic intervals and top/base-depth tables. On 60 validation wells, BG-UT achieves an average accuracy of 82.09% and a macro F1-score of 84.89%, exceeding SegNet and Encoder Transformer by 4.46 and 2.71 percentage points in average accuracy, respectively. The proposed method effectively improves boundary localization stability, reduces the influence of local abnormal log responses and stratigraphic thickness variations, and provides a reliable intelligent approach for fine-scale stratigraphic correlation and subsequent reservoir modeling under complex reservoir conditions.