The September 2026 issue of SPE's Journal of Petroleum Technology runs a technology focus on reservoir surveillance, and the interesting thing about it is not the framing essay. It is the roster. Three operator-authored papers sit under it, with a fourth listed alongside, and between them they cover most of the ways a mature field is watched.
PETRONAS reports time-lapse saturation surveillance by pulsed neutron logging on a thin-oil-rim brownfield in Malaysia, feeding continuous live model updating and the placement of new horizontal wells (OTC 36701). bp reports proactive surveillance and targeted data acquisition on the mature ACG field in the Caspian, feeding infill, completion and decline-management decisions (SPE 230423). Petrobras fuses tracer, production and injection data through fuzzy logic into an interwell connectivity index on a giant presalt field offshore Brazil (SPE 231704). Equinor reports tracer analysis at Breidablikk (SPE 231912). The framing essay is by a senior reservoir engineering specialist at ADNOC, who cites an IEA analysis of around 15,000 fields implying roughly 8% annual decline, about 5.5 million B/D a year, without continued investment.
These are first-party accounts: operator engineers reporting their own results, which is why they carry method where a press release carries adjectives. They are also, every one of them, an inference from a measurement made at or along a wellbore to a decision made about a volume of rock. That step is the one worth examining, because we have measured what it costs when it is taken on the wrong quantity.
The question that comes before accuracy
A surveillance programme is usually argued in terms of the instrument: how precise the tool is, how many wells carry it, how often it is run. None of those is the first question.
A measurement made in a wellbore speaks for the rock around it only as far as the property itself stays correlated. That distance is the lateral range of the property's variogram, and it belongs to the geology, not to the tool. A perfectly accurate measurement of a quantity with no lateral structure tells you about the rock the wellbore touched and about nothing else, however many wells you instrument and however good the instrument is.
We know what that looks like on real data, because we measured it before interpolating anything. On a 13-well Middle East carbonate field, our Detection Transformer produced four candidate curves from borehole image logs, and we ran every one through inter-well comparison before kriging. Bedding density correlated cleanly between wells. Fracture density did not: it was localised, a near-wellbore property rather than a layer-cake one. Bedding-plane strike yielded no significant inter-well correlation either, and fracture area proved a poor correlation identifier, useful per well and meaningless field-wide.
Three of four. Not marginal, not noisy: no measured lateral structure at all.
- Informed area, pattern cells
- 8.6
- If it extrapolated
- 25.4
- Share of well envelope
- 100%
- Lateral range
- 80 m or more
| curve | what we measured | lateral range | informed area here |
|---|---|---|---|
| bedding density (shown) | correlated cleanly between wells | bound, about 80 m or more | 8.6 pattern cells |
| fracture density | did not correlate, a near-wellbore property | none measured | 0 pattern cells |
| bedding-plane strike | no significant inter-well correlation | none measured | 0 pattern cells |
| fracture area | a poor correlation identifier, useful per well only | none measured | 0 pattern cells |
What a zero does to a well count
Select fracture density in the exhibit and drag the well slider from 1 to 13. Nothing rises. Select bedding-plane strike, then fracture area, and drag again. Nothing rises there either, under either interpolation rule, at any spacing in the band.
This is the finding, and it is a structural one rather than a caution. Informed area is the union of the supports around each instrumented well, so it grows with the square of the lateral range and only with the count of wells. When the range is zero the union is empty, and 13 wells multiply an empty set into an empty set. There is no well count, no acquisition budget and no improvement in tool precision that converts a non-correlating quantity into a field-scale one. Our case study named interpolating one anyway as the most dangerous failure mode of a correlation tool, because the output looks exactly like the output that was earned.
Now select bedding density, the one curve that earned it. At all 13 wells the informed area is 8.6 pattern cells, one cell being the inter-well spacing squared, and it covers the well envelope. Switch off the rule that kriging interpolates rather than extrapolates and the same field reads 25.4 pattern cells. Nothing was measured differently. The larger number is what you get by allowing an estimator to speak past the wells that constrain it, and it is the number to distrust when a surveillance programme is presented as covering a field.
One more thing about that curve, said plainly because the alternative is to overstate it. Our record reports the inter-well spacing as roughly 40 to 80 m and says the variogram range comfortably spanned neighbouring wells. That supports a lower bound of about 80 m and no more. We do not have a printed range for bedding density on that field, so the exhibit carries it as a bound and labels it as one everywhere it appears.
Why this is not the piece we published in September
On 9 September we published a piece that also put a correlation length on an axis, and it is close enough that the overlap belongs on the page rather than in a footnote. That one is about the semivariogram of a velocity model's residual at the closure scale, and its finding is a continuous trade: an estimator that cuts its RMS error while shortening the correlation length of what remains can improve its published score and make the relief error worse. The subject is an error field, and the lesson is that a scalar accuracy is a marginal statistic.
This one is the variogram of the rock property itself, and its finding is not a trade at all. It is a degenerate case. The question is not how far an error stays correlated but which quantities have any lateral structure to interpolate, and the answer on our field was that most of them did not. The two pieces meet at the same practical instruction, which is to ask for a variogram rather than a scalar, and then diverge: September's asks it of the residual, this one asks it of the property, before a single estimate is made.
What to ask of a surveillance programme
The four programmes above are not doing anything wrong. Tracer analysis and time-lapse saturation logging are measurements whose lateral reach is a real, and in several cases a long, physical thing, which is exactly why they were chosen. The papers are member-locked, so we are reading their synopses and we cannot see whether any of them prints a measured range. That is the gap worth closing, and it is closeable by whoever owns the data.
Three questions follow, and they are cheap.
Ask which quantity is being interpolated, not which tool measured it. A programme that reports tool accuracy has answered a different question from the one that decides how much rock it speaks for.
Ask for the range, and for the evidence behind it, as a measured value or an honest bound. A bound is a legitimate answer. A silence is not.
Ask what happens outside the well envelope. An estimator that interpolates is constrained by the wells; one that extrapolates is generating structure, and on our field the difference between those two answers was a factor of about three on the same data.
The pattern cell is the right unit for all of this, incidentally, because it is the one that survives contact with a real field. A percentage of a field flatters whoever quotes it and depends on where the licence boundary happens to be. A multiple of the pattern the wells actually form does not.
Key takeaways
- Four operator-authored papers in one JPT issue, from PETRONAS, bp, Petrobras and Equinor, all infer from a wellbore or path measurement to a field-scale decision.
- A measurement speaks for the rock around it only as far as the property stays correlated, and that range belongs to the geology rather than to the tool.
- On our own 13-well carbonate field, three of the four automated curves we produced had no measured inter-well correlation: fracture density, bedding-plane strike and fracture area.
- A curve with a lateral range of zero informs an area of zero at every well count, so no acquisition budget converts it into a field-scale quantity.
- For the one curve that did correlate, enforcing that kriging interpolates rather than extrapolates changed the informed area from 25.4 pattern cells to 8.6 on the same data.
- Ask which quantity is being interpolated, ask for its range as a value or an honest bound, and ask what the estimator does outside the well envelope.
References
- Reservoir Surveillance technology focus, Journal of Petroleum Technology, 1 September 2026, introducing OTC 36701 (PETRONAS), SPE 230423 (bp) and SPE 231704 (Petrobras), and listing SPE 231912 (Equinor). jpt.spe.org. Paper bodies are SPE member-locked; this piece uses the published synopses only.
- Our own delivery record on which curves correlate and which do not: Well-to-well kriging and the bedding-density correlation pipeline.
- The companion argument about an error field rather than a rock property: The error that cancels.
- N. Cressie. The Origins of Kriging. Mathematical Geology, 22(3), 239 to 252, 1990. doi:10.1007/BF00889887
