Abstract
An automated fracture picker is usually presented as the deliverable: feed it a borehole image, get sinusoids back with dip and azimuth. That framing quietly assumes the pick is the product. In a fractured carbonate it is not. The reservoir flows through its fractures, and the fractures carry flow through their aperture and their density, not through the matrix porosity a conventional log measures. This survey reads the published petrophysics that lives between a detection and a decision, and uses those numbers to place the AI pick where it belongs, at the front of a value chain that only pays off at the far end. The single-fracture relation reported on the Marun field ties permeability to the square of aperture, which is why a picker that resolves width to a few microns and one that does not can report permeabilities an order of magnitude apart. The Luthi-Souhaite aperture equation and its permeability index, validated over a 1 to 1000 md range against formation-tester, mud-loss and production data, mark the step where the physics gets fragile. And the imaging tool bounds everything upstream, since a fracture the sensor never illuminated cannot be picked, corrected, or turned into permeability. The claim is narrow: detection accuracy is necessary but not sufficient, because the number that changes a completion is a permeability, and permeability is an aperture problem.
Why porosity is the wrong axis
The instinct carried over from clastic reservoirs is that permeability tracks porosity, so a good porosity log is most of the answer. In a fractured carbonate that instinct fails. The matrix can be tight while the well produces, because the flow is in the fracture network, and a fracture contributes permeability through the cube of its aperture in the parallel-plate idealisation, reduced in the field literature to a single expression tying the permeability of one fracture to the square of its width. The Marun study, worked in the mixed carbonate-and-sandstone Asmari Formation, writes that reduction as the relation below, with the aperture w in inches, and pairs it with a connectivity constant C = 250/80 that sets how the individual fractures assemble into a producing network [1].
The 80 to 400 md figure that same study reports is a useful reference point but a different quantity: it is the matrix permeability of the study's clean sandstone zones, not a fractured-carbonate flow band. We carry it here only as an order-of-magnitude anchor for a permeable clastic matrix, the very thing a tight carbonate matrix is not, so that the contrast between a matrix number and a fracture-driven one stays legible. The number an engineer perforates on in this reservoir is set by the fracture population, not by that sandstone band.
The exponent is the whole story. Because K rises as the square of aperture, a width the front of the chain resolves to within a few microns and a width it smears out by tens of microns produce permeabilities that differ by an order of magnitude, and there is no porosity number that recovers the difference after the fact.
This is why a fracture picker in a carbonate is not measured by whether it drew a sinusoid in the right place. It is measured by whether the picked features, once counted per metre and corrected into apertures, reproduce the permeability profile the well actually flows. That is a much longer chain than a detection score, and each link can be strong while the chain is weak.
The links between a pick and a permeability
Reading the petrophysics forward, four things have to happen after a pick before an engineer has a number to act on. First, the individual picks become a density: fractures per metre, resolved along depth into a curve, because a single fracture rarely sets the flow but a swarm does, and the connectivity constant only means something once you know how many fractures there are per interval [1]. Second, each fracture gets an aperture. The standard route is the Luthi-Souhaite relation, which reads aperture from the excess conductivity a conductive fracture adds to the image, with tool-specific coefficients and the mud and flushed-zone resistivities [2].
Third, density and aperture combine into a permeability profile, and Aghli and colleagues validate exactly this: a permeability index built from image-log fracture analysis, checked over a 1 to 1000 md range against formation-tester points, mud-loss events and production [2]. Fourth, and only fourth, the profile informs a completion decision, where to perforate and where to stimulate.
The aperture step is the fragile one. It is fragile in two independent ways at once, and both matter precisely because of the squared relation upstream: an aperture that is off feeds a permeability that is off much harder. We do not re-derive that fragility here, because it is the whole subject of a companion piece, The Aperture Trap, which shows why the Luthi-Souhaite estimate runs away in high-resistivity-contrast rock and why image logs inflate any fracture thinner than 0.1 mm, and argues for reporting density and conductive-or-resistive class in place of a per-fracture width. For the value chain the point is only structural: this is the link where the physics is least trustworthy, so a picker that hands its output straight to an aperture routine, without carrying the fracture density it can report far more reliably, is optimising the wrong link.
What the imaging tool decides before the model runs
There is a link before the pick, too, and it is the one no model can repair. A fracture the sensor never illuminated cannot be detected, cannot be aperture-corrected, and cannot enter the permeability profile. Aghli and colleagues make the coverage difference concrete: an FMI tool carries 192 electrode buttons and illuminates roughly 80 percent of the borehole wall, while an EMI configuration carries 150 buttons for roughly 60 percent [2]. The 20-point gap is not a cosmetic difference in image quality. It is a fifth of the wall, and the fractures in that fifth are simply absent from every downstream step. A picker trained and scored on well-covered FMI images can post a strong detection number and still miss a real share of the fracture population on a more sparsely instrumented log, and the miss propagates all the way to the permeability the engineer trusts.
This is the sense in which the pick sits at the front and not at the centre. Everything above it, the tool and its coverage, sets the ceiling on what detection can recover, and everything below it, density, aperture, permeability, decides whether a good detection becomes a useful number.
Even the front of the chain is contested
None of this is an argument that detection is easy or solved. The competing literature makes clear that stage one is an active problem in its own right, particularly away from the vertical wells where most image-log methods were tuned. A conditional GAN approach reported by Wei in 2020, reviewed in the Swin-transformer segmentation work of Wang and Zhou, reaches about 90 percent accuracy specifically on horizontal and low-angle fractures, the geometries that give conventional sinusoid fitting the most trouble [3]. That result is worth taking seriously on its own terms. But it also reinforces the survey's point rather than undercutting it: a 90 percent detection accuracy on a hard fracture geometry is a strong front-of-chain number, and it still says nothing about whether the resulting density and aperture reproduce the permeability the well flows. The detection benchmark and the permeability benchmark are different benchmarks, and a program that reports only the first has left the reader to assume the second.
Discussion
Read together, these published numbers describe a value chain with a specific shape. The economically meaningful output is at the far end, a permeability profile precise enough to move a perforation or a stimulation call. The physics that connects the front to the end is unforgiving in one direction, because the squared aperture relation amplifies error rather than averaging it out, and fragile in one place, the sub-0.1 mm aperture regime where image logs are known to exaggerate. The imaging tool caps what is recoverable before the model sees a pixel. And detection, the part that gets demonstrated in a slide, is real work but is only the first of five links.
Where our own work sits relative to this map is a matter of emphasis. When we build a picker for a carbonate, the accuracy of the sinusoid is the entry ticket, not the prize; the outputs we treat as first-class are the ones that survive the chain, a fracture density that can be reported with confidence and an aperture handled with the caveats the literature insists on rather than as a headline width. The published petrophysics is not a constraint we chafe against. It is the specification for what the pick is for. An automated fracture detector earns its place when it shortens the path from a scanned image to a defensible permeability, and it is judged, in the end, by a number it does not itself produce.
Limitations
This is a survey of external published literature and inherits that literature's boundaries. The relation K = 5e7 w2 and the C = 250/80 connectivity constant are reported in the Marun study and are field-specific; the coefficient should not be transported to another reservoir without recalibration, and we use it here to argue the shape of the dependence, not to predict any particular reservoir [1]. The 80 to 400 md figure is the matrix permeability of that study's clean sandstone zones, not a fractured-carbonate flow band, and we carry it only as a reference for a permeable clastic matrix; nothing in the argument depends on treating it as the permeability of a fracture network. The Marun relation is published with the aperture w in inches; the instrument's plotted magnitudes follow that convention, and where the exact unit handling is approximated the curve is labelled illustrative rather than presented as a field-calibrated permeability. The Luthi-Souhaite aperture equation and the 1 to 1000 md validated index range come from a single carbonate study whose authors are explicit that the aperture estimate is unreliable in high-resistivity-contrast zones and inflated below 0.1 mm; the mechanics of that fragility are treated in the companion piece The Aperture Trap rather than re-argued here [2]. The FMI and EMI coverage figures are nominal tool descriptions, not a measured detection gap on a common well, so the claim that missing coverage propagates to permeability is a physical argument, not a controlled experiment. The competitor accuracy we cite is a single reported figure on a specific fracture geometry and is not a like-for-like comparison against the other detectors named [3]. Finally, the instrument's aperture sweep and its five-stage chain are an illustrative rendering of the sourced relation and the published workflow; only the coefficient, the sandstone reference band, the index range and the tool coverage are sourced numbers, and the plotted curve is the physics, not a fit to any field's measured permeability.
References
[1] Marun, M. Estimating the reservoir permeability and fracture density using image logs in the Asmari Formation, reporting the single-fracture crack-permeability relation K = 5e7 w2 (K in md, w the aperture in inches) and a connectivity constant C = 250/80 in the mixed carbonate-and-sandstone Asmari Formation; the 80 to 400 md figure is the matrix permeability of the clean sandstone zones, not a fractured-carbonate band. Petroleum Science and Technology 31(10): 1048-1056, 2013. https://www.tandfonline.com/doi/abs/10.1080/10916466.2010.540617
[2] Aghli, G., Moussavi-Harami, R., and Mortazavi, S. Fracture analysis and estimation of permeability index in carbonate reservoirs using image logs. Petroleum Science 17: 51-69, 2020. Applies the Luthi-Souhaite aperture equation, validates a permeability index over a 1 to 1000 md range against MDT, mud-loss and production, notes the sub-0.1 mm exaggeration caveat, and contrasts FMI (192 buttons, about 80 percent coverage) with EMI (150 buttons, about 60 percent). https://link.springer.com/article/10.1007/s12182-019-00413-0
[3] Wang, H., and Zhou, Y. A dual encoder-decoder Swin transformer for fracture and bedding segmentation in image logs, reviewing a conditional GAN (Wei 2020) that reports 90 percent accuracy on horizontal and low-angle fractures. Petrophysics 64(1): 38-49, 2023. https://onepetro.org/petrophysics/article/64/01/38/517537
