In August 2026 Saudi Aramco and Ma'aden signed a shareholders' agreement for a minerals exploration joint venture, 51 percent Ma'aden and 49 percent Aramco, over Zone-4 of the Arabian Platform: roughly 182,000 square kilometres, about a tenth of the Kingdom's land area [1]. Aramco's stated contribution is technical rather than financial. It is AI, high performance computing, subsurface modelling, and the geological and geophysical record of ninety years of petroleum exploration in one basin [2][3].
The trade press has read that record as a target-finding asset, and one outlet framed the whole venture as a bet that ninety years of data can find Saudi copper [2][4]. The record supports a narrower claim. The narrower claim is also the more useful one, because it can be tested.
What a petroleum corpus is actually made of
Reflection seismic works because sedimentary basins are layered and the layers differ in acoustic impedance. Every product built on top of it inherits that assumption: horizons, faults, sequence boundaries, amplitude anomalies, velocity models, and every neural network any of us has trained on migrated volumes. The corpus encodes where impedance changes across sub-horizontal surfaces, and it encodes it very well, because ninety years of drilling has tied those surfaces to rock.
Hard-rock copper and rare earth targets do not answer to that description. They sit in crystalline basement, and they are compact bodies, not the continuous layering the petroleum section is made of. Compactness on its own is not what hides them from seismic. Acoustic impedance is density times velocity, and massive sulphide is dense enough that most sulphide ores plot in a distinct impedance field, well clear of common silicate host rocks [10]. Reflection methods can be aimed at such a body directly, and surveys built for that job have imaged known ore bodies in central and eastern Canada [11].
The constraint is this corpus, not the method. A hard-rock survey aimed at a compact body is designed for that target: fold, station spacing and bandwidth chosen for energy scattered off an object a few hundred metres across in a crystalline host, and processing that keeps that energy instead of muting it as noise. Ninety years of Arabian Platform seismic was designed for the opposite job. It was shot, processed and tied for specular reflection off sub-horizontal sedimentary surfaces, with basement as the floor of the section and everything below it left where it lies.
That is the end of the target-identification story for these volumes. No relabelling exercise and no fine-tune turns them into a copper detector, because they were never acquired to record the class of object in question, and the well control that makes them trustworthy stops at the basement top. A model can only read what its inputs carry.
Where the corpus does transfer, and transfers well
Under cover, nobody measures the target. What a gravity or magnetic survey records is the sum of the target and everything above it, and everything above it is larger. Relief on the basement top of a few hundred metres, at a sediment-to-basement density contrast near 500 kg/m3, produces a gravity signature of several milligals. A large massive sulphide body under that cover produces a fraction of one. Stripping the overburden is the measurement. The anomaly hunt is what is left over afterwards.
Ninety years of seismic and well control is the best constraint anyone could hand you on exactly that stripping term: how thick the cover is, what it weighs, and how its base is shaped. This is a real contribution and it is worth paying for. It is also the only term the corpus enters.
The arithmetic that follows is short. The peak gravity anomaly of a compact body of radius $a$ and density contrast $\Delta\rho$ buried at depth $z$, together with the half-width of that anomaly, behaves like this.
The signal falls as the inverse square of burial depth and the anomaly widens linearly at the same time, so a deeper body is smaller and blurrier at once. The overburden term moves the other way. Treated as a Bouguer slab over basement relief $\delta h$, it grows with the cover you are trying to see through.
The two terms want opposite cover thicknesses
Both statements are about the same variable, cover thickness, and they pull it in opposite directions. The corpus is worth more where the cover is thicker, because there is more to strip. The target is worth less where the cover is thicker, because it is further away. The region where both conditions hold is an interval, not a half-line, and an interval has a width you can quote.
The exhibit below makes that width the readout. It carries a cutaway block of the transition zone, sedimentary cover over crystalline basement with a compact target seated in the basement, and two draped response surfaces above it: the target's own anomaly and the overburden response that has to come off first. The strip chart on the left plots two indices against cover thickness. One is target detectability against a three-sigma screening threshold. The other is the error the corpus removes, measured against the survey noise floor, because a correction smaller than the noise floor changes nothing anyone can see.
The two curves cross. Drag the cover slider and the cursor sweeps across a fixed pair of curves, carrying the two readings in opposite directions: target detectability falls, corpus gain climbs. On the default parameters the band runs from 60 metres of cover, below which the corpus is buying an accuracy nobody can detect, to 415 metres, above which the target drops under the threshold. That is 355 metres of cover, out of a slider that runs to 1,500. Turn the corpus off and the band falls to 140 metres. The 215 metres of cover the corpus adds is the size of the prize, and it is worth having. Push the survey noise floor to 0.2 mGal and the band closes completely. Grow the target to a 300 metre radius and it opens to 955 metres, which is the defensible version of the case for this venture: it works for large bodies under moderate cover.
Every number the model assumes is printed on the exhibit's own face, in the method note under the legend, down to the magnetic noise floor and the gravity floor it is quoted against. The only unprinted constant is Newton's. They are stated assumptions about screening physics, not survey results, and a reader who disagrees can see precisely which number to argue with.
The channel that works best under cover gets nothing from the corpus
Magnetics is the standard tool for basement targets beneath sedimentary cover, and the reason is the same term. Sedimentary cover is effectively non-magnetic, so there is almost no overburden response to strip out of a magnetic survey. That immunity is exactly why a petroleum corpus contributes nothing there. The corpus's whole contribution was to the stripping term, and where there is no stripping term the contribution is zero.
The magnetic channel pays for that immunity with a steeper depth law, the inverse cube rather than the inverse square. It survives deeper than gravity does in the model, and when it goes, it goes faster. Toggling the corpus control in the exhibit moves the gravity readout and leaves the magnetic readout exactly where it was.
What regime transfer has cost us, measured on our own wells
This is not a theoretical worry for us, and we do not have to argue it from first principles. Our own published borehole work shows the same collapse twice, on far smaller regime steps than the one being proposed here.
A combined bedding-and-fracture model on a Middle East carbonate operator's borehole-image logs reached a headline fracture F1 near 82 percent at a 9 cm depth tolerance on vertical wells. Run blind on a horizontal well it had never trained near, the same model family plateaued near 63 percent [8]. The cause was geometric and predictable. A fracture crossing a near-vertical borehole traces a tall sinusoid averaging about 75 cm on the unrolled image; the same fracture crossing a near-horizontal borehole traces a flat one averaging about 15 cm. Roughly a fifth of the amplitude, in the same field of view, and nineteen points of F1.
The second case is worse and more instructive. On the same engagement, a fracture configuration scored an F1 of 50.87 percent on its validation split at a 2 cm depth match and 3.69 percent on a well the model had never touched [9]. A factor of about fourteen. The validation number was not wrong arithmetic. It was measuring a different regime, because overlapping image patches had leaked across a random split, and the blind well removed the leak by construction.
Regime distance is not a discount factor you apply to a score. It changes which physical quantity the model is allowed to read. Vertical to horizontal inside one field is a small step and it cost nineteen points. Sedimentary basin to crystalline basement, acoustic impedance to density and susceptibility, petroleum to minerals, is not a small step. Anyone quoting a transfer number across it is quoting a hope.
How this compares with what the other majors have claimed
Equinor reported that artificial intelligence saved it around USD 130 million in 2025, with the gains coming from its own operations and its own data [5]. ADNOC and AIQ announced an agentic AI system for energy operations, built inside the operating regime it was designed to run [6]. ExxonMobil said it had identified four exploration opportunities offshore Guyana with AI, in a petroleum basin, from petroleum data, using petroleum physics [7].
Read the three together and the pattern is that every one of them applies a model inside the regime its data came from. None is a cross-domain claim. The Aramco and Ma'aden venture is the first of these announcements to ask a petroleum corpus to work outside its own physics, which is what makes it worth writing about, and which is also why the defensible version of the pitch is narrower than the headline.
Four questions before believing a screening number
- Which term does the corpus enter? If the answer is target identification, ask which of those volumes were acquired over crystalline basement, at what fold and bandwidth, and what ties them below the basement top.
- What is the cover-thickness distribution across the licence, and what fraction of the area falls inside the band where both conditions hold?
- What is the survey noise floor on each channel, and does the corpus-derived correction exceed it? A correction below the noise floor is real and undetectable at the same time.
- What is the blind result, on holes the model never saw, in the same cover regime as the ones it will be asked to rank?
The licence area is 182,000 square kilometres. The screening advantage is not defined over 182,000 square kilometres. It is defined over the fraction whose cover thickness sits inside the band, and nobody outside the venture knows what that fraction is, ourselves included. Where the Platform thins westward toward the exposed Shield, the corpus has almost nothing to strip and a regional gravity grid gives you the same answer. Where it thickens eastward, the target is gone before the corpus becomes worth having. In between there is real work for ninety years of seismic and well control, and it should be done. It is a smaller claim than the one in the headlines, and unlike that one it can be checked.
Key takeaways
- Massive sulphide carries enough impedance contrast to be a seismic target where a survey is built for it; Aramco's volumes were shot, processed and tied for a layered sedimentary section, so they hold nothing about a compact body in crystalline basement and the corpus cannot transfer to the target term.
- Where it does transfer is the overburden-stripping term: cover thickness, density and basement-top structure, which is the largest single error in any potential-field screening under cover.
- Corpus value grows with cover thickness while target detectability falls as the inverse square of it, so the two conditions cross and the usable band is an interval a few hundred metres of cover wide, not a half-line.
- Magnetics, the tool of choice under sedimentary cover, carries almost no overburden term and therefore gains nothing at all from ninety years of seismic control.
- Our own published record puts numbers on regime transfer: 82 percent on vertical wells read 63 percent blind on horizontal ones, and a 50.87 percent validation F1 read 3.69 percent on a well the model had never touched.
References
[1] Saudi Aramco. Aramco and Ma'aden sign shareholders' agreement for minerals exploration joint venture (August 2026). https://www.aramco.com/en/news-media/news/2026/aramco-and-maaden-sign-shareholders-agreement
[2] Enterprise AM. Aramco is betting 90 years of data can unlock Saudi's copper potential (26 August 2026). https://enterpriseam.com/ksa/2026/08/26/aramco-is-betting-90-years-of-data-can-unlock-saudis-copper-potential/
[3] Engineering and Mining Journal. Aramco and Ma'aden form JV for Saudi Arabia (August 2026). https://www.e-mj.com/breaking-news/aramco-and-maaden-form-jv-for-saudi-arabia/
[4] Mining Outlook. Aramco and Ma'aden form mining JV to target copper and critical minerals in Saudi Arabia (August 2026). https://www.mining-outlook.com/regions/middle-east/saudi-arabia-mining/aramco-and-maaden-form-mining-jv-to-target-copper-and-critical-minerals-in-saudi-arabia
[5] Equinor. Artificial intelligence saved Equinor USD 130 million (7 January 2026). https://www.equinor.com/news/20260107-artificial-intelligence-saved-equinor-usd-130-million
[6] ADNOC. ADNOC and AIQ developing first-of-a-kind agentic AI solution for global energy transformation (2024). https://www.adnoc.ae/en/news-and-media/press-releases/2024/adnoc-and-aiq-developing-first-of-a-kind-agentic-ai-solution-for-global-energy-transformation
[7] World Oil. ExxonMobil identifies four Guyana exploration opportunities using AI (13 August 2026). https://www.worldoil.com/news/2026/8/13/exxonmobil-identifies-four-guyana-exploration-opportunities-using-ai/
[8] EarthScan. From 82% on paper to 63% blind: an honest accounting of the vertical-to-horizontal generalization gap. https://earthscan.io/insights/from-82-on-paper-to-63-blind-vertical-to-horizontal-generalization-gap
[9] EarthScan. Validation said 51%, the blind well said 4%: the error ledger that kept us honest. https://earthscan.io/insights/validation-said-51-the-blind-well-said-4-the-error-ledger-that-kept-us-honest
[10] Salisbury, M. H., Milkereit, B., and Bleeker, W. Seismic imaging of massive sulfide deposits, Part I: rock properties. Economic Geology, 91(5), 821-828 (1996). Laboratory density and velocity measurements at elevated pressure place massive sulphide ores in a distinct acoustic impedance field, separated from common silicate rocks. doi:10.2113/gsecongeo.91.5.821
[11] Salisbury, M. H., Milkereit, B., Ascough, G., Adair, R., Matthews, L., Schmitt, D. R., Mwenifumbo, J., Eaton, D. W., and Wu, J. Physical properties and seismic imaging of massive sulfides. Geophysics, 65(6), 1882-1889 (2000). Predicts sulphide acoustic impedance from modal mineralogy and reports 1D, 2D and 3D seismic experiments over known ore bodies in central and eastern Canada. doi:10.1190/1.1444872
