A speed claim is the first thing a reader believes and the first thing a reviewer attacks. It is one number, easy to quote in a slide and easy to get wrong in a way that costs you credibility on everything else in the paper. We learned this the direct way. An earlier draft of the carbonate work carried a processing figure of about 30 seconds per metre for the vug-quantification pipeline, stated twice, and a reviewer asked us to withdraw it because the comparison it rested on was not fair to the methods it was measured against. We took the number out. The lesson was not that the pipeline is slow. It was that the 30-second figure was the wrong number, attached to the wrong model, defended against the wrong baseline.
This piece is about the number that is right. For the fracture and bed detection model, built on an end-to-end detection transformer, the defensible figure is under 10 seconds to interpret a 5 metre section. That is what we put on the World Petroleum Technology Congress poster, and it is the claim we can stand behind in review. Below is the honest speed math: what the 10 seconds covers, what it leaves out, and what it is faster than.
Two models, two speed numbers, one of them retired
The confusion starts because the programme produced two separate detectors on the same carbonate image logs, and they are not interchangeable. One finds fractures and beds by regressing sinusoid parameters end to end. The other quantifies vugs, the secondary-porosity cavities, with a classical adaptive-thresholding and contour pipeline. Different mechanics, different outputs, different speed behaviour.
The retired 30-seconds-per-metre figure belonged to the vug pipeline. It was withdrawn for a specific reason: the reviewer showed that comparing that vug-processing time against fracture-detection methods was comparing two things that do not do the same work. A vug quantifier counting individual pores every 0.1 m and a fracture detector fitting sinusoids are not racing on the same track, so the time ratio between them says nothing useful. Once the comparison is unfair, the number has to go, whether or not the pipeline is fast. So the vug speed claim came off the manuscript.
None of that touches the fracture and bed model. Its speed number was measured on its own work against its own baselines, and it survives the same scrutiny that killed the other one.
What the 10 seconds actually covers
The claim is under 10 seconds to interpret a 5 metre section. Be precise about the boundary of that measurement, because a speed number is only honest if you say where it starts and stops.
The 10 seconds is inference time for the trained detection transformer over a 5 metre interval: taking the preprocessed, dynamically normalised image patches, passing them through the network, and emitting the set of detected sinusoids with their depth, dip, and azimuth parameters. It is the cost of turning pixels into picks once the model exists and the data is prepared.
It does not include model training. Training is a one-off before any log is interpreted, and folding it into a per-section time would be dishonest arithmetic, because one trained model serves every subsequent section. It does not include the data-preparation steps ahead of inference: loading the log from its binary file, converting the missing-pad sentinel values to blanks, and applying the dynamic normalisation. Those are real costs, but they are shared setup, not per-section interpretation. And the 10 seconds is a speed statement, not an accuracy statement. It says how fast the picks come out, not how right they are; pick accuracy lives in a separate set of depth, dip, and azimuth threshold metrics.
State the boundary that plainly and the claim stops being a marketing figure and becomes a measurement. A reviewer can disagree with where we drew the line, but they can see the line.
What it is faster than
A speed number means nothing without the thing it beats. The fracture and bed model is measured against two published references, and the gap to each is the substance of the claim.
The first is the semi-automatic prior art. I2AM, from Ferraretti and colleagues, interprets petroleum-geology image logs and is a fair comparison because it targets the same kind of data. Its cascade extracts machine-vision features, clusters them, and trains a classifier on the result. The reported benefit is that it cuts interpretation time from days to hours. That is a real improvement over fully manual work, but note where it stops: a geologist still has to cut the clustering dendrogram to choose the facies. That human step holds it in the hours class. It never reaches seconds, because it was never meant to run without a person. [1]
The second is a classical-computer-vision baseline. Li and colleagues published a path-morphology method for extracting features from image logs at roughly 5 minutes per metre. [2] For a 5 metre section that is about 25 minutes of processing, against under 10 seconds for the transformer on the same length of log. The gap is not marginal. It is the difference between a method you run interactively and a method you run overnight.
The exhibit puts all of this on one axis. The manual expert read sits in the hours; the semi-automatic prior art sits in its days-to-hours band with a human still cutting the dendrogram; the classical baseline sits at 5 minutes per metre; and the published claim sits low on the ladder at under 10 seconds for 5 metres. Drag the section length and the argument holds: the published figure scales cleanly with length while the manual and prior-art regimes balloon, because they carry per-metre human cost the end-to-end model does not. The withdrawn 30-seconds-per-metre vugs figure is tagged off the ladder on purpose, a different model's number kept out of a comparison it was never entitled to join.
Why end to end is what buys the seconds
The gap comes down to where the human sits. The semi-automatic cascade needs a geologist to cut the dendrogram before it classifies anything, so its wall-clock time is bounded below by human availability, not compute. The classical path-morphology method has no human in the middle but pays a heavy per-metre compute cost sweeping morphological operators across the image. The transformer removes both: it regresses the sinusoid parameters directly, treating each planar feature as an object with a depth, dip, and azimuth to predict, and once trained it is a single forward pass over the interval. That is why the number is small, and why it stays small as the section grows.
The number is the discipline, not the headline
The throughput and cost consequences of interpreting a well this fast, weeks of expert backlog collapsing to hours, are their own argument and we have made it elsewhere; this piece is only about whether the speed number itself is defensible. The honest version of a speed claim is not the biggest number you can produce. It is the number you measured on the right model, against a fair baseline, with the boundary of the measurement stated out loud. The 30-seconds-per-metre vug figure failed that test and we retired it. The under-10-seconds-for-5-metres fracture and bed figure passes it, which is why it is on the poster and the other one is not.
Limitations
The under-10-seconds figure is an inference-time claim for the fracture and bed model over a 5 metre section and excludes model training and the shared data-preparation steps ahead of inference; a full end-to-end wall-clock time including load and normalisation would be larger. It is a speed statement only and reports nothing about pick accuracy, which is governed by separate depth, dip, and azimuth threshold metrics. The comparison baselines are drawn from their authors' reported figures on their own data and hardware, not re-run on ours, so the ladder compares stated speeds across studies rather than a controlled head-to-head on identical inputs. The manual expert read shown on the exhibit is an illustrative hours-class anchor for the phrase "versus manual hours," not a measured figure from this engagement. All results come from a carbonate reservoir in Oman and other rock types or logging conditions may shift the numbers.
References
[1] Ferraretti, D., Gamberoni, G., and Lamma, E. I2AM: a semi-automatic system for data interpretation in petroleum geology. A semi-automatic image-log interpretation cascade that reduces interpretation time from days to hours while keeping a geologist in the loop to cut the clustering dendrogram. https://www.springer.com/journal/10844
[2] Li, B., and co-authors. Automatic fracture-vug identification and extraction from electric imaging logging based on path morphology. A classical path-morphology feature-extraction method reported at roughly 5 minutes of processing per metre of log. https://doi.org/10.1016/j.petrol.2019.106508
[3] Nasim, M. Q., and co-authors. Deep learning for enhanced fracture and bed detection in subsurface geological analysis. Conference poster, World Petroleum Technology Congress, 19 April 2024, stating the interpretation-speed claim of under 10 seconds for a 5 metre section. https://www.17wptc.com/
