Who Is Talking to Whom Underground: The Identifiability Problem in Injection-Production AI

CNOOC Limited's interim results of 26 August 2026 say that "the intelligent injection-production interaction scenario for offshore oilfield production was selected as a high-value scenario at the 2026 World Artificial Intelligence Conference" [1]. That is the only item on the page that is specifically artificial intelligence rather than digital infrastructure, and it is the one this piece is about, because injection-production interaction is the hardest estimation problem on a mature waterflood and the least visible from outside.

Two weeks earlier, on 11 August 2026, Baker Hughes announced a multi-year contract with Kuwait Oil Company for the Ahmadi Innovation Valley project, and said that the technologies "will benefit from Baker Hughes' portfolio of digital and artificial intelligence (AI) automation solutions that help operators increase recovery from existing wells, lower operating costs, reduce water production and minimize power consumption" [2]. On 30 June 2026 SLB announced a seven-year contract under the same Ahmadi Innovation Valley initiative, covering artificial intelligence, industrial internet of things applications, production optimization, reservoir technologies, water management and energy transition [3]. Sinopec's interim release of 23 August 2026 reports the launch of what it calls "the industry's first digital expert", the Fenghuo industrial AI agent, and separately, under sci-tech development, that "breakthroughs were made in synergistic oil flooding theories and intelligent drilling methods" [4].

Four releases, three operators, and one distinction that decides how to read all of them.

Who is speaking, and about what

CNOOC is the operator speaking about its own scenario. Sinopec is the operator speaking about its own agent, and separately about its own oil flooding research: the flooding sentence sits under sci-tech development and is not described as AI anywhere on the page, so nobody should read it as an AI result, including us [4]. Both are first-party, and both are statements of intent rather than of deployment. A scenario selected at a conference is a scenario selected at a conference.

The Kuwait Oil Company legs are different in kind, and they differ from each other. The Baker Hughes announcement quotes only its own chief executive, Lorenzo Simonelli, on delivering "tailored technology solutions at scale that improve production performance and efficiency, supporting KOC's goals to maximize value from their assets" [2]. No KOC executive is quoted in it, and the sentence about reducing water production and increasing recovery from existing wells is Baker Hughes describing its own portfolio, not KOC describing its own field [2]. The SLB release does quote KOC's chief executive, Ahmad Jaber Al-Eidan, who says that Ahmadi Innovation Valley "represents an important step in advancing technology leadership across Kuwait's energy sector" and that "through collaboration with leading technology partners, we are accelerating technology deployment, strengthening local capabilities and expanding knowledge transfer to support Kuwait's energy industry" [3]. That is the operator speaking, first-party, and what it is speaking about is a programme. The list that contains water management, reservoir technologies and production optimization is SLB's own sentence describing what SLB will do under the agreement, and the nearly 100 projects the initiative spans is SLB's count [3]. Both are real awards from a real operator programme. Neither carries a KOC statement about what AI has done to a waterflood, and this piece does not treat them as though it did.

None of the four releases attaches a figure to the AI content. There is no recovery uplift, no water-cut reduction, no allocation accuracy, and no count of wells. That is not a criticism: three are results announcements and one is a contract award, and none of them claims to be a technical paper. It is the reason the rest of this piece is arithmetic rather than commentary. The capability claims are checkable against the releases. The question of whether the data behind them can identify what the models are being asked to identify is checkable against nothing but the design of the field's own operations.

What an injection-production model is actually estimating

Strip an intelligent injection-production system down to the quantity it exists to produce and you get a connectivity matrix: for each producer, the fraction of each injector's water that reaches it. That matrix is what a rate-allocation decision runs on. If injector three is feeding a watered-out producer and injector four is feeding a good one, you cut three and hold four. Every optimisation on top of the model, agentic or otherwise, is a function of those numbers.

The standard estimator for this from rate data alone is a capacitance-resistance model. In its steady-state form, a producer's liquid rate is a weighted sum of the injection rates around it,

$$q(t) \;=\; \sum_{i=1}^{M} f_i \, w_i(t) \;+\; \varepsilon(t), \qquad \varepsilon \sim \mathcal{N}(0, \sigma_q^2)$$

FormulaLaTeX

Copy LaTeX

so estimating the connectivity vector $f$ is a linear regression whose design matrix is the injection history itself. That is the sentence the whole argument turns on. The operator does not choose the design matrix in a laboratory. The design matrix is whatever the field did.

And what fields do is move their injectors together. A waterflood is run to a voidage replacement target, the injectors share facilities, pumps, water supply and power, and when the target moves they all move. Write the schedules as deviations about their own means, each with standard deviation $s_w$ and pairwise correlation $\rho$. Over $T$ samples the design's Gram matrix is

$$X^{\top}X \;=\; T\,s_w^2\left[(1-\rho)\,I \;+\; \rho\,\mathbf{1}\mathbf{1}^{\top}\right]$$

FormulaLaTeX

Copy LaTeX

With an isotropic prior of standard deviation $\tau$ on each allocation fraction, the posterior precision is that Gram matrix over the noise variance, plus the prior, plus whatever a designed perturbation adds. Every one of those terms is a diagonal plus a rank one:

$$P \;=\; \alpha I + \beta \mathbf{1}\mathbf{1}^{\top}, \qquad \alpha = W(1-\rho) + G + \tfrac{1}{\tau^{2}}, \qquad \beta = W\rho$$

FormulaLaTeX

Copy LaTeX

The two widths

A precision matrix of that shape has exactly two distinct eigenvalues, and they are the whole argument. Along the all-ones direction, the total support the producer receives, the eigenvalue is $\alpha + M\beta$. Along every contrast direction, which injector supplies that total, it is $\alpha$, with multiplicity $M-1$. So the posterior standard deviation on one allocation fraction and on the total are

$$\mathrm{sd}(f_1) = \sqrt{\frac{\alpha + (M-1)\beta}{\alpha\,(\alpha + M\beta)}}, \qquad \mathrm{sd}\!\left(\textstyle\sum_i f_i\right) = \sqrt{\frac{M}{\alpha + M\beta}}$$

FormulaLaTeX

Copy LaTeX

and the ratio of the worst-resolved direction to the best-resolved one is

$$A \;=\; \sqrt{\frac{\alpha + M\beta}{\alpha}}$$

FormulaLaTeX

Copy LaTeX

Now put more data through it, of the kind a field already has. Bottomhole pressure logged at the producers alongside rate is a second observation channel, and it is driven by the same injection schedules, so it multiplies $W$ and leaves $\rho$ alone. Both eigenvalues grow. But $\alpha + M\beta$ grows faster than $\alpha$ whenever $\rho > 0$, so the ratio goes up, and as the prior becomes negligible it approaches a limit that contains neither the noise nor the length of the record:

$$\lim A \;=\; \sqrt{\frac{1-\rho+M\rho}{1-\rho}}$$

FormulaLaTeX

Copy LaTeX

At five injectors and a schedule correlation of 0.90 that limit is the square root of 46, which is 6.78. No amount of passive measurement on that schedule goes past it, and every improvement in measurement moves toward it rather than away from it. The only term in the precision that changes it is $G$, and $G$ is what you get when somebody changes injection on purpose, on a schedule chosen so the injectors do not move together.

The bench

The exhibit below computes that surface rather than drawing it. Left to right is the correlation between the injector schedules, from 0 to 0.99. Back to front is rate-measurement noise, from 1% to 20% of the producer rate. Height is the posterior standard deviation on one allocation fraction, with the ceiling at the prior of 0.200. The amber sheet is an allocation tolerance of 0.050, which is roughly the resolution needed to tell a 25% connection from a 15% one, and the bright line is where the surface crosses it. The white curtain is your correlation, and the labels follow it.

The toy is synthetic and the plate says so. Five injectors on one producer, a steady-state capacitance-resistance balance, 156 weekly rate samples, an injector schedule swing of 800 barrels a day about its own mean, a producer at 5,000 barrels a day, a pulse of the same amplitude as an ordinary swing, and a pressure channel worth one rate channel. There is no operator's field in it and no client data. Every one of those numbers is an assumption, which is why every one of them is either printed on the plate or sitting on a control.

correlation between injector schedules, 0.00 to 0.99: 0.90rate-measurement noise, 1.0% to 20.0%: 8.0%designed pulse testing, 0.0 to 104.0 weeks: 0.0bottomhole pressure at the producers: not logged

What the bench shows

At the opening settings, a schedule correlation of 0.90 and 8% rate noise with no designed test and no pressure channel, one allocation fraction is known to 0.096 against a prior of 0.200. The record has removed about half the width of a guess. The total support the producer receives is known to 0.042 over the same period, and the ambiguity ratio reads 5.76 times. The model knows how much water is arriving. It does not know which injector is sending it.

Now tick the bottomhole pressure box, which is the intervention every asset team reaches for first because the gauges are usually already in the well. The hero improves from 0.096 to 0.073 and the total from 0.042 to 0.029. The crossing label moves from a correlation of 0.50 to 0.76, so the tolerance now survives a much more correlated schedule. And the ambiguity ratio goes the other way, from 5.76 to 6.20, while the figure beside it, the ambiguity the design allows, does not move at all: it reads 6.78 with the box ticked and 6.78 with it clear. That is the finding, and it is reached by ticking one box and reading a number that gets worse while the two beside it get better.

Push the same lever harder and the gap closes rather than opens. Drag the noise slider to its 1.0% end with pressure logged and the hero falls to 0.010, a coefficient known to a hundredth of an allocation fraction, while the ambiguity ratio reads 6.77 against the 6.78 the design allows. Better metering buys precision all the way down and buys identifiability not at all. It walks the ratio up to the ceiling the schedule set and stops.

The pulse control is the one that moves the ceiling. From the opening state, drag it to 26 weeks: the hero falls to 0.065, the ambiguity ratio to 3.96, and the figure that would not move for pressure falls to 4.23. At 104 weeks the three read 0.041, 2.56 and 2.62. The readout on the right prints what the tolerance actually costs: from the opening state, 60.2 weeks of designed testing brings one coefficient to 0.050, and with the pressure channel already logged, 43.6 weeks does it. Those two numbers are the procurement question. They are also the reason the pressure channel is worth having, once you stop asking it to do the job it cannot do.

Two more places on the surface are worth visiting. Drag the correlation to 0.99, which is what a field running hard to a single voidage target looks like: the hero reads 0.160 against a prior of 0.200, the record has learned almost nothing about any individual injector, and the design allows an ambiguity of 22.27 times. Tick pressure there and the ratio climbs from 9.99 to 12.88. Then drag the correlation to 0.00, a field where every injector moves independently: the hero reads 0.039, the ambiguity ratio reads 1.00, and the pulse readout says the designed test is not needed. But look at the total in the left readout, which reads 0.088 at a correlation of 0.00 against 0.042 at 0.90. Correlated schedules are not simply bad. They excite the common mode hard and they resolve the total support better than an uncorrelated schedule does. What they destroy is the contrast, which is the only thing an allocation decision uses.

What our own record does and does not cover

We should be exact about where this sits relative to work we have published. We have no public case study on ensemble history matching or on connectivity estimation, and the posterior widths above come from a synthetic toy, not from an engagement. The statement that the posterior stays wide across injectors until schedules are decorrelated is a statement about our own unpublished runs and about the algebra above, and it is offered as such.

What we have published that bears on it is an illustrative scenario of the surrounding machinery, a composite case study rather than a client record. In it, an agentic MLOps loop across roughly 40 producing assets has four agents monitor drift, reconcile data, retrain and validate, and production-forecast accuracy lifts 22% across the portfolio as models stay current [5]. The part of that scenario most relevant here is the failure it describes: one early retrain passes every statistical test and violates basic reservoir physics on two wells, which is why the loop carries a mandatory physics gate that checks material balance and keeps predicted permeabilities inside reservoir-analog bounds before a geoscientist is asked to approve anything [5].

A physics gate catches a model whose output is impossible. It does not catch a model whose output is merely unidentified: a connectivity vector that honours material balance exactly, forecasts the producer's rate exactly, and puts the water in the wrong place, because the data cannot tell that arrangement from the right one. On a correlated schedule the posterior is nearly flat along that whole family. Any of them fits. The one the optimiser hands the asset team is set by the prior, the regulariser and the random seed, and no validation gate we have written would flag it, because there is nothing wrong with it except that it is not identified.

Three questions before a waterflood AI programme

What is the correlation between your injector schedules over the history the model will be trained on? It is a five-minute calculation on data every operator already has, and it is the coordinate that decides which regime the programme is in. Below a modest correlation the ordinary record identifies the connectivity and the work is engineering. Above it, no model class fixes the problem, because the problem is in the design matrix and not in the estimator.

What would you have to add for the model to be identified, and is it a sensor or a decision? On this bench a second measurement channel narrows every width and widens the ambiguity, while a designed perturbation is the only control that moves the ceiling. That is a claim about a synthetic toy, but the mechanism is the eigenstructure of an equicorrelated design, and that mechanism does not care whose field it is.

How many weeks of deliberately decorrelated injection is the asset willing to run, and what does that cost in deferred production? That is a trade the reservoir engineers can price and the AI vendor cannot. The releases at the top of this piece describe capabilities: an interaction scenario, an automation portfolio, a research valley, a flooding theory. Not one of them describes the injection design behind it, and until somebody does, the honest thing to say about an intelligent injection-production model is that it knows the total.

Key takeaways

1. 1CNOOC Limited says its intelligent injection-production interaction scenario was selected as a high-value scenario at the 2026 World Artificial Intelligence Conference, and attaches no figure to it. Sinopec names the Fenghuo AI agent and, separately and not as AI, breakthroughs in synergistic oil flooding theories. On the Kuwait Oil Company legs, Baker Hughes and SLB describe their own scope under the Ahmadi Innovation Valley programme; no KOC executive is quoted in the Baker Hughes release, and the KOC chief executive quoted in the SLB release speaks about technology deployment and knowledge transfer, not about a waterflood.
2. 2Estimating injector-to-producer connectivity from rate history is a linear regression whose design matrix is the injection schedule itself. The operator does not choose that design in a laboratory; a waterflood run to a voidage target moves its injectors together, and that correlation is what decides whether the connectivity is identified.
3. 3An equicorrelated design gives a posterior with two widths: narrow along the total support a producer receives, wide along which injector supplies it. On a synthetic five-injector toy at a correlation of 0.90 and 8% rate noise, one allocation fraction is known to 0.096 against a prior of 0.200 while the total is known to 0.042, an ambiguity of 5.76 times.
4. 4Logging bottomhole pressure at the producers improves both widths, 0.096 to 0.073 and 0.042 to 0.029, and makes the ambiguity worse, 5.76 to 6.20. Better metering does the same: at 1.0% noise with pressure logged the coefficient reaches 0.010 and the ambiguity reaches 6.77, against the 6.78 the schedule alone allows. More of the same data climbs toward that ceiling and never past it.
5. 5A designed pulse campaign is the only control on the bench that moves the ceiling: 26 weeks takes the ambiguity from 5.76 to 3.96 and the figure the design allows from 6.78 to 4.23. Reaching an allocation tolerance of 0.050 from the opening state costs 60.2 weeks of designed testing, or 43.6 weeks with the pressure channel already logged.
6. 6We have no published case study on ensemble history matching or connectivity estimation, and the toy is synthetic. What we have published is an illustrative agentic MLOps scenario, in which the loop lifts production-forecast accuracy 22% across roughly 40 assets and a physics gate is added after a retrain violates reservoir physics on two wells. A physics gate catches an impossible model. It does not catch an unidentified one.

References

[1] CNOOC Limited. CNOOC Limited Focuses on Value Creation, Production and Profit Hit New Highs in H1 2026. 26 August 2026. https://www.cnoocltd.com/english/presscenter/pressreleases/2026/202608/t20260826_122439.html

[2] Baker Hughes. Baker Hughes Awarded Multi-Year Contract by Kuwait Oil Company for Ahmadi Innovation Valley Project. GlobeNewswire, 11 August 2026. https://www.globenewswire.com/news-release/2026/08/11/3342544/0/en/baker-hughes-awarded-multi-year-contract-by-kuwait-oil-company-for-ahmadi-innovation-valley-project.html The release is Baker Hughes' account of its own contract and portfolio; no Kuwait Oil Company executive is quoted in it.

[3] SLB. SLB Awarded Seven-Year Contract Under Kuwait Oil Company's Ahmadi Innovation Valley Initiative. Published 30 June 2026. https://www.slb.com/newsroom/press-release/2026/pr-2026-0630-slb-koc-aiv This is SLB's own release; the quotation attributed above to Ahmad Jaber Al-Eidan is the Kuwait Oil Company chief executive speaking inside it, and the scope list and the count of nearly 100 projects are SLB's own sentences. World Oil reported the same award, worldoil.com, and is trade press rather than either company's own account.

[4] China Petroleum and Chemical Corporation (Sinopec Corp). Press Release: Sinopec FY2026 Interim Results. EQS Newswire, 23 August 2026. https://www.eqs-news.com/news/corporate-news/en-press-release-sinopec-fy2026-interim-results/514ef82e-238a-4aa2-9f2d-32f1161d68b0_en The oil flooding sentence appears under sci-tech development and is not described as artificial intelligence on the page.

[5] EarthScan. Illustrative scenario: Agentic MLOps: From 6-week retrains to overnight. A composite case study, not a client engagement. https://earthscan.io/case-studies/agentic-mlops-six-week-retrains-to-overnight