A conference poster is a strange forcing function. You have one A0 sheet and a few minutes of a passer-by's attention, and into that you have to fit work that took a small team the better part of two years: the ingestion, the quality control, the model, the metric, and the evidence that each design choice earned its place. In April 2024 we did exactly that for the fracture-and-bed detection work at the World Petroleum Technology Congress, a venue our earlier conference recaps never touched. This is the story of what survives that compression, and of two ablation numbers that lived only on the poster.
Why the congress mattered for a research partnership
The work itself is not new to anyone who has followed this programme. It is a transformer-based detector, built on the DETR family, that reads an unwrapped borehole image log and returns the sinusoids a planar feature traces across the borehole wall, each as a triple of depth, dip, and azimuth. What the congress added was reach. A peer-reviewed paper reaches reviewers and citation graphs; a poster at a large petroleum-technology venue reaches the people who actually run image-log interpretation for operators, standing in front of the sheet asking whether the numbers hold. For a partnership between a research lab, an academic institution, and a major operator in Oman, that face-to-face credibility beat is not a side effect. It is part of the point, and it sits alongside the Omanization commitment the programme carried from the start, training young Omani professionals into the same subsurface-AI work the poster describes.
Condensing the pipeline onto one sheet also forces an honest question about what the argument actually is. You cannot fit every table. You keep the QC that proves the data was trustworthy, the model that does the reading, and the ablations that prove the design was not arbitrary. Everything else is supporting text.
The QC that earns the top third of the sheet
The first thing on the poster is not the model. It is the evidence that the input was worth modelling. Two checks carried their own figures.
The first catches the logging tool lying about its own range. A dynamic image channel should span a wide intensity range; on one delivery the static channel arrived quantised into a 0-to-15 band where we expected something closer to 0-to-255. A pixel-range histogram on the poster makes that anomaly visible at a glance, because a model trained on a silently compressed channel learns the compression, not the rock. The second check is a sampling-interval plot that flags depth bands where the logging run skipped intervals, the kind of gap an expert scanning by eye can miss and a training set cannot afford to inherit. Both checks answer the same reviewer question before it is asked: how do you know the data under the model is clean?
The model, in one equation
The detector's job is to turn a patch of the unwrapped image into a set of sinusoids. Each sinusoid a feature traces is a curve
and the network regresses the parameters that map to the three quantities an interpreter cares about, depth, dip, and azimuth. Training uses bipartite matching between predicted and ground-truth sinusoids through a Hungarian assignment, a focal loss on the presence classification, and an absolute-error loss on the parameters. The 800-pixel patch height, roughly 2.2 metres, comes straight from the measured distribution of sinusoid heights, so that more than 95 percent of features fit inside a single patch. None of that is new to this programme. What the poster added, and what makes it worth a case study rather than a reprint, is that it carried its own ablation evidence.
The two numbers that lived only on the poster
The ablations are how a poster answers "why this and not that" without a conversation. We ran two on the poster whose figures differ from the journal-paper set, and the difference is the interesting part.
The first strips the model of the dynamic image and feeds it the static channel instead. Dynamic and static are two normalisations of the same tool response; dynamic stretches contrast locally so faint features stay legible, static preserves absolute scale. On the poster's numbers, static matching runs a class error of 70.36 percent against dynamic's 3.59, with the Hungarian matching loss moving from 0.15 to 0.03 and the parameter loss from 0.56 to 0.08. The second turns off the augmentation that expands each labelled patch into many. Without it the class error is 100 percent, a model that never learns the feature at all; with it, 3.62, with the loss pair falling from 0.25 and 0.65 to 0.05 and 0.09.
Those are not the journal figures. The peer-reviewed set reports 63.45 versus 2.536 for the matching ablation and 100 versus 2.618 for augmentation, both covered elsewhere in this series and not re-derived here (see "Static vs Dynamic Image Logs" and the well-count and augmentation write-ups). The poster carried a different snapshot of the same experiments, taken at a different point in the model's life. Keeping the two sets distinct is a small discipline that matters: a number on a poster is a claim about a specific run, and quoting the journal value under a poster figure would be quietly wrong.
Read the exhibit either way and the shape is identical. The arm we argue against pushes class error into the 70-to-100 band; the arm we ship holds single digits. That is what an ablation on a poster is for. It is not a full study, it is a decisive contrast a reader can absorb in the time it takes to read two bars.
The speed line we have since narrowed
The poster also carried a headline we would now write more carefully. It claimed the model reads a 5-metre section in under 10 seconds, a substantiated speed line at the time and a genuine improvement over the multi-minute-per-metre conventional methods the field was used to. We have since narrowed how we talk about that number. A wall-clock reading from a tool log is a property of a specific machine, batch size, and section, not a portable benchmark, and an earlier vugs-quantification claim in the same programme was withdrawn for exactly this reason. The honest version is that the served model is fast enough that interpretation stops being the bottleneck, and the precise figure belongs to the run that produced it, not to a marketing sheet. We keep the poster's number here because it is what the poster said; we would not put it on a new one without the qualifier.
What condensing a pipeline to one sheet is worth
The transferable asset is not the poster. It is the compression exercise. Forcing a multi-year pipeline onto one sheet is a good audit of whether you understand your own work: it makes you name the three things that carry the argument, the QC that earns trust in the data, the model that does the reading, and the ablations that prove the design, and it makes you throw out everything that was only ever scaffolding. It also surfaces the small integrity choices that are easy to skip in a long paper. Which run does this number come from. Is this speed claim portable or is it a wall-clock reading. Does the poster's ablation set match the journal's, and if not, which one am I quoting. A venue that puts you in front of the people who run the tools for a living rewards getting those right, and it is a cheaper way to earn an operator's trust than another revision cycle.
Limitations
The numbers here are from a single confidential carbonate programme in Oman and the specific runs behind one conference poster; they are not a benchmark to quote against a different tool stack or dataset. The poster ablation figures differ from the peer-reviewed set for the same experiments, and both are snapshots rather than a converged final result. The sub-10-seconds-per-5-metres speed line is a wall-clock reading tied to a particular machine and batch configuration, which is why we treat it as a since-narrowed claim rather than a portable figure. The mechanics of the matching, well-count, and augmentation ablations are covered in the companion pieces named below and are not re-derived here.
References
-
World Petroleum Technology Congress poster on deep learning for fracture and bed detection in subsurface geological analysis, presented April 2024; poster-variant ablation figures (static-vs-dynamic 70.36 vs 3.59; augmentation 100 vs 3.62), QC figures, and the sub-10-seconds-per-5-metres speed line derived from the congress poster archive; data and code withheld under operator confidentiality.
-
Companion write-ups in this series covering the peer-reviewed ablation set, including the static-versus-dynamic image-log comparison and the well-count and augmentation ablations, which report the journal figures (63.45 vs 2.536 and 100 vs 2.618) not re-derived here.
