Acoustic borehole images are sonar pointed at geology. A rotating transducer descends through the borehole, bouncing sound off the surrounding rock. The reflected signal's amplitude and travel time encode the mechanical properties of the formation — hard rock reflects strongly, soft rock absorbs, fractures scatter. The result is a 360-degree acoustic image of the borehole wall, unrolled into a flat strip.
Interpreting these images requires segmenting them into geological units: this band is limestone, that band is shale, the bright spot is a fracture. Expert interpretation is expensive and subjective. Automated segmentation requires labeled training data that doesn't exist at scale.
A weakly supervised approach bypasses the annotation bottleneck (arXiv:2603.20729). Threshold-based pseudo-labels — automatically generated from simple rules applied to the image statistics — provide initial training targets. A neural network learns to segment using these rough labels, then refines its predictions beyond what the thresholds could capture alone. The key innovation is a confidence-gated depth-aware cross-attention mechanism that fuses the two-dimensional acoustic image with one-dimensional well-log data recorded at the same depths.
The cross-attention is what matters. The acoustic image has spatial texture — patterns that vary around the borehole. The well-log data has depth trends — properties that change with formation. Neither alone is sufficient: the image misses depth context, the well-log misses lateral variation. The gated fusion allows each modality to inform the other where it's confident and defer where it's not.
The structural insight: the underground is known through two complementary probes that measure the same rock differently. The acoustic image captures spatial variation that the well-log misses. The well-log captures depth trends that the image misses. Neither is the “true” picture — the geology is the structure that explains both measurements simultaneously. Understanding the subsurface is not choosing between modalities but finding the interpretation that is consistent with both.