We were hired to detect fractures, beddings and vugs on borehole image logs, and we did. But every well we ran through the interpretation kept handing us a second dataset we were not asked to build: the well's stress state, written on the same image in the same geometry, free of charge. The vendor's interpretation report for one carbonate well says it plainly. The borehole breakouts sit on a mirror azimuth running roughly E-W. The drilling-induced fractures run NNE-SSW. And the natural fractures in that interval are enhanced by the induced ones. Those two azimuths are not incidental features. They are the horizontal-stress orientation of the reservoir, the exact quantity a geomechanics study exists to recover.
This piece is about a door that was scoped, closed, and left ajar. The stress signal is already in the picks. The picker that finds it is one detection class away.
The signal that came free with the interpretation
Start with the physics, because it is why the azimuths matter. When a well is drilled, the rock around the borehole feels the far-field stresses concentrated at the wall. Where the wall rock fails in compression, it spalls into a pair of breakouts on opposite sides of the hole, and that breakout axis points along the minimum horizontal stress. Where the wall rock fails in tension, thin drilling-induced fractures open, and those run along the maximum horizontal stress, orthogonal to the breakouts [1]. Read the two off an image log and you have fixed the horizontal-stress orientation without a single lab measurement.
That is precisely what the interpretation report for this well contains. Breakouts on an E-W mirror axis. Induced fractures on NNE-SSW. The two families sit close to orthogonal, as the mechanics demands. The report also notes that the natural fractures are enhanced by the induced set, which is a real interpretive judgement about the reservoir and, separately, a labelling headache we will come back to. Alongside the stress signal, the same interpretation carried the structural picture: a gentle structural dip of 4.2 degrees toward 288.2 degrees, roughly WNW. None of this required extra logging. It fell out of reading the image.
What the picker already produces, per well
Now the machine-learning side. Our detection stack does not read stress today. It reads fractures and beddings, and it reads a lot of them. On this one well the fracture census came back sorted into the four standard image-log classes: 1 continuous conductive fracture (CCF), 335 discontinuous conductive fractures (DCF), 3 continuous resistive fractures (CRF), and 67 discontinuous resistive fractures (DRF), for 406 fractures in a single interval, plus 910 bedding dips. Every one of those is a sinusoid the detector localised, parameterised, and handed back as a depth-dip-azimuth triple. The apparatus that produces that census is a Detection Transformer trained to emit a fixed set of object queries, each regressing one feature's sinusoid.
The load-bearing observation is this: a borehole breakout and a drilling-induced fracture are not exotic new objects to a model that already detects fractures. They are features on the same unrolled azimuth-by-depth raster, distinguished by their appearance and their orientation rather than by a different sensor. Natural fractures cross the borehole and draw full sinusoids; induced fractures and breakouts run closer to parallel with the borehole axis and draw a different, recognisable signature. Adding them is a new detection class, not a new detection problem. The census above is the proof that the hard part, turning a raw image into localised, parameterised features at production volume, is already solved.
The door that 2020 scoped and the project closed
The reason this matters is not academic. Geomechanics was the original ambition. The 2020 proposal that started this whole line of work was, on paper, a geomechanics project, not the fracture-and-vug project that actually got built. Its first phase planned a mineralogy-based brittleness index. Its second phase planned a neural network to solve five unknowns, Young's modulus, Poisson's ratio, the two horizontal stresses, and the borehole size, from two known caliper diameters, calibrated against core. That was the technical core the engagement was meant to have.
It was descoped. The executed programme became detection transformers for fractures and beddings, vug computer vision, and well-to-well correlation, all of it valuable, none of it the stress-field analysis the proposal opened with. The brittleness index and the five-from-two network never got built. The door was scoped and then shut.
The instrument above is the argument in one frame. The compass on the left plots the stress signal that is already sitting in the interpretation: the E-W breakout mirror axis, the NNE-SSW induced fractures, the natural fractures the induced set enhances, and the white pointer for the 4.2-degree structural dip toward 288.2 degrees. The ladder on the right shows the four fracture classes the picker already ships, the 406 fractures and 910 bedding dips per well, and the one rung that is missing. Add the breakout-and-induced-fracture class, the orange rung, and the geomechanics box lights up: the 2020 scope, reactivated, because the azimuths it needs are exactly the ones the new class would detect.
Why "one class away" is a claim, not a slogan
It would be easy to overstate this, so let me be precise about what is and is not implied. Adding a detection class is genuine work. It needs labelled examples of breakouts and induced fractures, and the natural-versus-induced ambiguity the report itself flags means those labels are not trivial to draw cleanly, since an induced fracture can look like an enhanced natural one. The published analogue is real but limited: a fast-region convolutional network reported roughly 98% AUC on fractures and 90% on breakouts, and it did so on simulated, single-event acoustic images, not multi-feature production microresistivity logs [2]. That gap between a clean synthetic benchmark and a messy real interval is exactly where the honest cost of the new class lives.
What "one class away" does mean is concrete. The pipeline that would host the new class already exists and already runs at well scale, as the 406-plus-910 census attests. The training target, a set of parameterised sinusoids, is the same target the picker already regresses. And the downstream payoff is not speculative: two orthogonal stress azimuths per well are enough to orient a horizontal-stress field, which is the first input the descoped brittleness-and-stress network was designed to consume. The distance from where we are to where the 2020 proposal wanted to be is one detection class, one set of labels, and the stress signal is already in the interpreted wells waiting to be read by a model instead of by eye.
That is the roadmap item worth writing down. Not a new project, but a class added to an existing one, opening a door the labels were already holding ajar.
Limitations
The stress azimuths, the fracture census, the 910 bedding dips, and the 4.2-degree structural dip are sourced from the interpretation report for a single carbonate well and the original proposal; they are not a multi-well stress inversion, and one well's orientation should not be generalised across a field without well-to-well work. The breakout-and-induced-fracture detection class described here is a roadmap item, not a shipped result: no such model has been trained on these logs, and the natural-versus-induced ambiguity noted in the report is a genuine labelling risk that could depress accuracy relative to the clean fracture classes. The external accuracy figures in reference [2] are from simulated acoustic imagery and single-event benchmarks, so they bound the analogue optimistically rather than predicting performance on production microresistivity data. Finally, orienting a stress field is a necessary but not sufficient step toward the full 2020 geomechanics scope, which also required caliper-derived magnitudes and core calibration that fall outside what a detection class alone can supply.
References
[1] Zoback, M. D. Reservoir Geomechanics. Cambridge University Press (2007). The reference treatment of how borehole breakouts and drilling-induced tensile fractures on image logs record horizontal-stress orientation.
[2] Dias, L. O., et al. Automatic detection of fractures and breakouts patterns in acoustic borehole image logs using fast-region convolutional neural networks. Journal of Petroleum Science and Engineering 191 (2020), 107099. AUC around 98% for fractures and 90% for breakouts on simulated acoustic images.
