Automated Geomechanical Interpretation of Borehole Images Using Context-Aware AI Mechanism

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Automated segmentation of ultrasonic borehole image logs is a critical step in formation evaluation and geomechanical characterization. Borehole images are inherently sequential: geological features such as breakouts, natural fractures, and stylolites extend continuously across depth intervals that may span tens to hundreds of meters. Yet the computa-tional necessity of dividing these long images into independent fixed-size patches severs this depth continuity, causing standard patch-based segmentation models to produce fragmented and topologically inconsistent predictions at patch boundaries. This paper addresses the geological continuity problem in borehole image segmentation by introducing a memory-augmented SegFormer (SWM) in which a dedicated Memory Attention Mechanism propagates contextual information across consecutive depth patches. Inspired by video object segmentation, the mechanism treats the ordered patch sequence as analogous to a video stream. It comprises three components: a Memory Encoder that fuses encoder feature maps with prior segmentation masks into compact tokens using ConvNeXt blocks; a circular Memory Bank storing the four most recent patch representations; and a Memory Attention Module with four stacked layers of self-attention and cross-attention using Axial Rotary Positional Encoding. A learnable gated fusion mechanism controls how much historical context the decoder receives at each depth interval. The approach is validated on borehole breakout segmentation as a representative use case – a task where longitudinal continuity is geologically fundamental. Evaluated on 10 wells acquired with SLB TerraSphere and UBI tools, SWM achieves average improvements of +3.36% in IoU and +3.88% in clDice over a standalone SegFormer baseline, with a false-negative rate reduction of 4.54%, confirming that explicit inter-patch memory substantially restores prediction continuity.