In a deep geothermal programme in Denmark, ADR predicted formation-boundary depths to within a few 10s of metres of the drilled wireline (over 2000m of subsurface depth) and tracked the temperature gradient closely — across six sites, blind-tested against real wells. It’s a clear example of how surface-acquired ADR turns into subsurface intelligence you can plan a drill programme around.
Classifying a 4 km geothermal well from surface — and its reservoir sands
A blind-tested ADR machine-learning lithology model over two deep Copenhagen Basin wells, Denmark — validated against the drill logs.
See before you drill.
A deep geothermal well lives or dies on its reservoir sandstone: get the depth, thickness and quality of the target sands right, and the well flows; get them wrong, and a multi-million-euro hole disappoints. VEXRAD tested ADR against that question over two deep wells in the Copenhagen Basin, Denmark. From surface, ADR built a ~4 km virtual borehole at each well, and a machine-learning lithology classifier was trained and blind-tested against the wells’ own drill-log interpretation. The result was a genuine validation: in blind cross-validation the ADR sand prediction correlated with actual drill-log sand content at r = 0.89, the model reproduced the full chalk–sand–silt–shale column, and it resolved the target reservoir sands directly — then carried the interpretation more than a kilometre below the deepest log.
A virtual borehole, then a blind test
ADR returns a depth profile — a “virtual borehole” — of the subsurface’s material-property response, acquired from surface. At each of two deep Copenhagen Basin wells, that profile was reduced to a set of ADR parameters (energy, frequency, dielectric response) and fed to a Random-Forest lithology classifier, trained on the well’s log-derived lithology (chalk, sand, silt, shale) and then blind-tested — a 5-fold cross-validation that repeatedly holds out a fifth of the labelled log and predicts it from data the model never saw. That is the honest test of a remote method: not whether it fits the log it learned, but whether it predicts the parts it didn’t.
Against the drill logs
Findings are tagged with the tier they earn — confirmed against the log, expected from the physics, or an untested target.
| Finding | Evidence | Tier |
|---|---|---|
| Full-column lithology | The ADR classifier reproduced the well’s logged chalk–sand–silt–shale column across ~4 km (CV Macro-F1 0.62) | Validated |
| Blind sand prediction | 5-fold cross-validation: 89% sand recall, and the ADR P(Sand) profile correlated with actual drill-log sand content at r = 0.89 | Validated |
| Reservoir sands (S6 & S7) | Resolved against the logs at ~2,462–2,552 m and ~2,622–2,642 m (VSAND 43–63%) — priority reservoir targets | Validated |
| Beyond total depth | ADR lithology prediction continues ~1,342 m below the deepest log, to ~4,020 m — reconnaissance-level | Grounded |
| Primary anomaly at 2,377 m | A strong ADR event, cross-confirmed at both wells and still undrilled — the programme’s priority new target | Candidate |
The sands that matter
For a geothermal well the reservoir is everything. ADR resolved the two principal target sandstones directly against the well’s lithology log — S6 at roughly 2,462–2,552 m and S7 at roughly 2,622–2,642 m, both with high sand fractions (VSAND 43–63%). These are the intervals a geothermal developer most needs to locate and characterise before committing the well.
What this doesn’t claim
- Validation is against the wells’ log-derived lithology (CPI), not physical core — a strong reference, but an interpretation itself.
- The classifier is probabilistic: several thinner sands were partial matches, including transition beds, and the sand false-alarm rate was around 58%.
- A sub-5 m bed falls below ADR’s ~5 m sampling and was retained as a manual pick, not an independent ADR detection.
- Predictions below the deepest log (to ~4 km) are reconnaissance-level — useful for context, not a substitute for drilling.
We report these plainly. That is the point of the four-tier framework.
What it means for geothermal developers
A deep geothermal well is a large, largely irreversible bet on a reservoir you cannot see. This work shows ADR reading the reservoir sands from surface — their depth, thickness and sand quality — blind-tested against real well logs at two sites, with a strong correlation to actual sand content, and extending the picture below where any log reaches. Calibrated on a well or two in a basin, ADR becomes a low-cost way to de-risk the next target before the rig arrives.
Talk to us about your project