From weeks to hours per well: the throughput dividend of automated interpretation

The manual baseline, costed honestly

A high-resolution borehole image log is unwrapped into a 2D image; the interpreter scans it, picks every fracture and bedding sinusoid, and records depth, dip, and azimuth for each. On this operator’s data, sinusoid picking ran to days per well, and vug identification took hours per well, both done by a scarce specialist.

That cost compounds. The engagement spanned roughly 16 wells of usable image-log data, against a broader inventory exceeding 80 image logs, so the manual process set the throughput ceiling for the entire subsurface programme. Manual picking is also inconsistent between interpreters and across sessions, so the multi-week spend bought a result that still carried inter-interpreter variance.

The reading that matters to a subsurface manager: this is a capacity problem disguised as a quality problem. The team did not lack skill; it lacked hours. Any intervention that did not move the throughput needle would have left the actual constraint untouched.

A production system, not a model

The instinct in subsurface AI is to chase the model. We built the model, a customised Detection Transformer, but the throughput dividend came from treating interpretation as a production system end to end, with the model as one component. The engineering had four load-bearing layers:

• The model. A DETR-derived detector with a ResNet-10 backbone and a multi-task head that classifies fracture-versus-bedding and regresses depth, dip, and azimuth in one forward pass. Set prediction with Hungarian matching recovers overlapping sinusoids natively rather than fighting them with thresholds.
• The data engineering. Reservoir intervals were brutally sparse. We grew the training corpus from roughly 900 image-ground-truth pairs to more than 55,000, a 65x increase, through overlapping patch extraction and geometry-preserving augmentation.
• The serving layer. Both tools were productised behind a containerised web application, so a geoscientist could run a well through interpretation from a browser, with no model code and no GPU babysitting.
• The MLOps and infrastructure. The whole thing ran on an on-prem GPU stack with experiment tracking, a versioned data store, and a custom MLOps layer, the scaffolding that lets a model be retrained, versioned, and trusted.

That fourth layer is where most subsurface-AI pilots die, and it is why the speed-up was real and repeatable rather than a benchmark artefact.

The throughput dividend, in numbers

With the pipeline in place, interpretation ran at roughly 30 seconds per metre of image log, about 5x faster than the manual baseline. The model picked sinusoids across multiple wells in minutes where the manual process took days per well, and the multi-week backlog collapsed into same-day turnaround.

Speed did not come at the expense of trust. On the productised stack the operator’s team reported +60% interpretation productivity and +75% interpretation accuracy, and the companion well-to-well correlation tool hit 95% target-location precision and 90% stratigraphic success, because a model applies the same picking criteria to every metre of every well.

Why this is an ROI story, not a research result

A 5x speed-up is not a vanity metric; it is a re-pricing of the constraint. Three commercial consequences followed directly.

• Same-day turnaround changes the operating model. When interpretation takes hours, not days, image-log analysis can sit closer to operations, informing completion and stimulation decisions inside a useful window rather than arriving after the fact.
• The expert-hours freed compound across the portfolio. Across a 16-well programme, against an inventory exceeding 80 image logs, the reclaimed specialist time is recurring capacity that grows with every new well.
• Consistency de-risks downstream decisions. A reservoir model built on machine-picked sinusoids inherits a single repeatable standard rather than the variance of several interpreters working under deadline.

The honest limits

The numbers are real but bounded. The headline figures are the operator’s own, measured on the productised tools over a confidential 16-well carbonate dataset; they are not a published, peer-reviewed benchmark. The corpus was vertical wells of a single reservoir; horizontal wells, where sinusoid amplitude drops sharply, are a separate distribution. And the speed-up assumes the production scaffolding stays in place: a model handed over as a pickle file would have decayed back into a research artefact within a quarter.

A model is one layer of four

Hover a layer. The speed-up came from the whole stack, not the model alone.

1. IV

   MLOps & infrastructure

   On-prem GPU stack, experiment tracking, versioned data store, custom MLOps.

   The scaffolding that lets a model be retrained, versioned, and trusted rather than re-derived by hand each time. This layer is why the speed-up was durable rather than a one-off.

   Where pilots die
2. III

   Serving layer

   Both tools productised behind a containerised web application.

   A geoscientist runs a well through interpretation from a browser, with no model code and no GPU babysitting.

   Browser-run
3. II

   Data engineering

   Grew the corpus from ~900 to >55,000 image-ground-truth pairs.

   Overlapping patch extraction and geometry-preserving augmentation, the difference between a model that learned nothing and one that converged.

   65x corpus
4. I

   The model

   A DETR-derived detector: ResNet-10 backbone, multi-task head.

   Set prediction with Hungarian matching recovers overlapping sinusoids natively, rather than fighting them with thresholds and non-maximum suppression.

   DETR