If you train a detection model on borehole image logs, you inherit a problem the geologists solved decades before you arrived: the image in front of you is not all rock. Some of what looks like a bed or a fracture is an artifact of how the tool ran, how the hole was drilled, or how the software rescaled the pixels. A trained interpreter carries a mental catalogue of these and discounts them by eye. A model has no such catalogue. It will fit a sinusoid to a tool-orbiting smear as happily as to a real fracture, and it will regress a dip and azimuth off a mis-oriented pass without ever signalling that anything is off. This is a note for the person building that model on a roughly twenty-month carbonate engagement with a mid-sized Middle East operator. The argument is narrow: artifact literacy is a prerequisite skill you gate on before training, not a caption you add after.
Four boxes that have not changed since 1999
The reference the interpreters use is old and still correct. Lofts and Bourke sorted borehole-image artifacts into four categories in a 1999 Geological Society special publication, and the taxonomy has held up because it is organised by cause rather than by appearance. The four boxes are acquisition-drilling, acquisition-logging, borehole-wall, and processing-derived.
Acquisition-drilling artifacts come from the bit and the trip: tool orbiting, bit-at-rest marks, wiper-trip scratching, side-track windows. Acquisition-logging artifacts come from the logging run itself: mud smear on the buttons, dead buttons or dead flaps, excessive tool rotation past one revolution per thirty feet, 60 Hz electrical noise, and faulty inclinometry. Borehole-wall artifacts are the shape of the hole imprinting on the image: washout, the key-seat furrow a wireline grinds into a deviated well, spiral hole, and drilling-induced fractures and breakout. Processing-derived artifacts are the ones the software invents: woodgrain, honeycomb gain error, tiger-striping, halo and proximity effects, and an incorrect hole diameter fed into the processing.
Learning these names is not trivia. It is the vocabulary that lets a team say out loud what it is looking at, and a model can only be gated against features its builders can name.
The two that lie to the labels
Here is where the taxonomy stops being a glossary and starts being an engineering constraint. Most of those artifacts are visible corruptions. Woodgrain looks like woodgrain. A washout blows out the image contrast in a way any reviewer catches. Tiger-striping stripes. You can see them, so you can decide what to do about them.
Two artifacts do not behave that way, and they sit in two different boxes. Faulty inclinometry is an acquisition-logging fault: the orientation channels that tell you how the tool was rotated and how the well was deviated are wrong, so the picture unrolls into a perfectly plausible image while the dip and azimuth computed from it are shifted. An incorrect hole diameter is a processing-derived fault: feed the wrong caliper into the image processing and the apparent dip comes out wrong, again with nothing visibly broken in the raster.
Those two numbers, dip and azimuth, are exactly the ground-truth channels our detection model regresses. We train a Detection Transformer to emit, for each planar feature, a depth, a dip, and an azimuth, matched against the interpreted picks. If the picks were computed off a bad inclinometry pass or a wrong diameter, the labels are wrong before a single gradient step, and the loss curve looks fine because the model is faithfully learning corrupted targets. The picture passed the eyeball test. The labels failed silently.
The guide above is the whole point in one frame. Sort the artifacts into their four canonical boxes and flag only the ones that poison the regressed labels, and two things become obvious at once. First, the dangerous artifacts are the minority. Lofts and Bourke say plainly that artifacts generally constitute the minority of a data-set, which is exactly why a team relaxes about them. Second, that minority does not cluster in one place you can guard. It is split across acquisition-logging and processing, and neither one announces itself in the image.
What a wrong pixel range actually cost us
The gate is not hypothetical. Early in the engagement we received ten wells and ran a value-range check on the raw static imagery before anything downstream. Two came back with abnormal static value ranges and were excluded on the spot, leaving eight usable, a fifth of the incoming data set removed at the door because the numbers were physically implausible.
The signature of a corrupt log is coarse and cheap to catch once you know the expected span. In the problematic well the compact-imager static values ran only 0 to 15, against the roughly -10^4 to 10^4 range a standard static image occupies. When that range collapse propagates into the static-to-dynamic conversion, it surfaces as unexplained amplitude spikes in the converted image, which is what we ended up documenting to a reviewer as conversion corruption rather than real signal. A pixel-range assertion at the front of the pipeline catches this before it becomes a training set.
The wells we cut for value-range failure and the artifacts that poison labels are the same problem wearing two coats. Both are cases where the image is not what it appears to be, and both are invisible to a model that was handed the data as trustworthy. The 1999 taxonomy is the older, richer version of the same discipline our value-range gate enforced: recognise the artifact before you interpret, or the interpretation inherits the fault.
A pre-training QC checklist
The practical output is a short list you run before a well enters a training set, in roughly this order.
- Assert the pixel range on the raw static and dynamic imagery. Physically implausible spans, such as a static image reading 0 to 15, disqualify the well until explained. This is the cheapest gate and it caught two of ten wells for us.
- Confirm the orientation channels are present and sane before trusting any dip or azimuth. Faulty inclinometry produces wrong dip and azimuth with a clean-looking image, so an absent or frozen tool-angle or well-deviation channel is a label hazard, not a cosmetic gap.
- Confirm the hole diameter that fed the processing is the real caliper. A wrong diameter yields a wrong apparent dip through no fault visible in the raster.
- Scan for the visible-corruption artifacts by name: washout, spiral hole, key-seat furrow, woodgrain, honeycomb, tiger-striping, dead buttons and flaps. Decide per zone whether to mask, keep, or drop, rather than letting the model fit them.
- Record which frame your labels are in. Apparent and true dip differ, and the correction depends on the same orientation channels the inclinometry check just validated.
None of these steps needs a model. They need a person who can name what they are looking at and a few lines of code that assert what the numbers should be. The artifact catalogue the interpreters have carried since 1999 is a data-quality specification for anyone training on these logs, and treating it as a gate is cheaper than discovering, after a clean-looking training run, that a fifth of your dip and azimuth labels were quietly wrong.
Limitations
This is a field guide, not a detector. The four-category taxonomy and every artifact name here come from Lofts and Bourke 1999; we applied it, we did not extend it. The value-range gate and the two-of-ten well exclusion are specific to one carbonate engagement and its mix of standard and compact microresistivity imagers, and the exact thresholds that count as an abnormal range are dataset-specific. The two artifacts we flag as label-poisoning are the ones that map onto the dip and azimuth channels our model regresses; a pipeline with different targets would draw the poison line elsewhere. Artifacts remain the minority of a data set, so the checklist filters the exception, not the common case.
References
[1] Lofts, J. C. and Bourke, L. T. (1999). The recognition of artefacts from acoustic and resistivity borehole imaging devices. In: Lovell, M. A. et al. (eds), Borehole Imaging: Applications and Case Histories. Geological Society, London, Special Publications, 159, 59-76.
