KAI-Subsurface · reference case on real public data
What core and logs tell you about permeability, and what they miss
Well Zeerijp-3A was cored over 200 m of Rotliegend sandstone in the Groningen gas field. We ran the KAI-Subsurface workflow on its public core analysis and logs: rock typing, permeability for uncored intervals tested blind, and the step from plug to bed. Every chart below is interactive.
Real data: well ZEERIJP-03-SIDETRACK1 (ZRP-3A), drilled by NAM in 2015, published on NLOG under the Dutch Mining Act. Only the bedding models in part 3 are simulated, and they are calibrated on these plugs.
Core plugs used409after quality control · 8 cores, 3,521–3,867 m
Rock types, where there is core96 %within a factor of 3 · one trend: 77 %
Logs only, tested blind60 %within a factor of 3 · 93 % within a factor of 10
kv/kh, plug against bed0.95 → 1 : 235median plug pair → 1 m bed with 30 % mud
1 · Where there is core
Three rock types predict permeability better than one trend
The flow zone indicator (Amaefule et al., 1993) groups plugs that share a pore geometry. Three types put 96 % of plugs within a factor of 3 of the measured value; one porosity–permeability trend manages 77 %. The tight, cemented type (RT1) is the one a single trend gets most wrong.
2 · Where there is no core
Permeability from logs, with an honest P10–P90
A random forest trained on the cores predicts permeability every 0.1 m from density, neutron, photoelectric, gamma-ray, resistivity and sonic logs. The band is the spread of the trees, widened to the error of the blind test. Grey intervals have log responses unlike any cored rock, so the prediction there is not trusted: 44 % of the logged interval is inside the cored range.
How the blind test works. Each of the eight cores is hidden in turn; the methods learn from the other seven and predict the hidden plugs from logs alone. Rock types predicted from logs do not help here (55 %; the right type 59 % of the time): logs read about half a metre of rock and cannot see the centimetre laminae that separate the types. That limit is why every KAI permeability log carries its band.
3 · From plug to bed
Plugs say kv/kh is about 1. The bed does not.
78 vertical–horizontal plug pairs give a median kv/kh of 0.95: at plug scale the sandstone looks isotropic, because plugs are cut in clean sandstone. Ripple-laminated bedding calibrated on the same plugs, and upscaled by flow simulation, gives kv/kh of 1 : 20 to 1 : 360 at 1 m for 20–40 % mud. Plugs cut where a laboratory cuts them overstate the bed's vertical permeability 20 to 1,200 times.
Build the bedding yourself in the Plug-to-Block Lab →Carry it between wells in the Heterogeneity Lab →
What is real and what is modelled
- Real: routine core analysis of about 600 plugs in 8 cores (porosity, Klinkenberg permeability, orientation and description) and the 2015 wireline logs of ZRP-3A, from NAM via NLOG.
- Real: the depth match between core and logs, found from the best correlation of core and log porosity.
- Set aside: fractured, broken or disintegrated plugs, which measure the damage rather than the rock.
- Modelled: the bedding in part 3, built from ripple templates whose sand and mud permeabilities come from these plugs.
- Not redistributed: NLOG files are free to download but carry no open licence; the notebooks fetch them from NLOG and check their SHA-256.
How it works
- Quality-check the plugs, then compute the flow zone indicator and group it into rock types (Gaussian mixture, number checked by BIC).
- Depth-match core to logs, then hide each core in turn and predict it from logs: one trend, rock types from logs, and a random forest.
- Train on all cores, predict every 0.1 m, give P10–P90 and flag rock unlike anything cored.
- Build bedding templates calibrated on the plugs, solve flow through them and drill virtual plugs (Nordahl and Ringrose, 2008; Ringrose et al., 2005).
Amaefule et al. (1993), SPE 26436 · Nordahl and Ringrose (2008), Math. Geosci. 40, 753–771 · Ringrose, Nordahl and Wen (2005), Petroleum Geoscience 11, 29–36.
Reproduce this case yourselfThree executed notebooks, the scripts that fetch the data from NLOG, and these results, open on GitHub (code MIT, text and results CC BY 4.0).Notebooks on GitHub ↗