Experimental and future technologies: Chapter A — Seafloor exploration · Chapter B — Minerals and the subsurface
Research Archive & Validation History
GeaSpirit fuses computational materials science with satellite remote sensing to detect mineral prospectivity from public-access data. This page is the engine's working logbook — schematic and honest: what each run measured, what failed, where the real ceiling is, and the road to a sellable exploration report and, ultimately, a mining-asset marketplace.
Operating philosophy. Make → Measure → Fix → Repeat. Every sprint must produce something demonstrable. Stack: 100% public-access data and open tools (Sentinel-2/1, Landsat, SRTM, EMIT, USGS / national geological surveys, Macrostrat, Google Earth Engine). No invented figures, ever.
1. The pipeline
For any area of interest the engine runs a fixed sequence and outputs a ranked target list with confidence:
- Acquire — pull public-access satellite, terrain, geophysics and geology layers over the AOI.
- Feature build — spectral mineral indices, thermal climatology, terrain/structure, geology, vegetation.
- Type-aware fusion — select the optimal feature subset for that deposit type and zone.
- Score & calibrate — a calibrated ranking surface under spatial cross-validation. It orders ground by model score; it is not a probability that mineralisation is present.
- Rank & report — ranked targets with exact coordinates, evidence breakdown and limitations.
2. Sprint plan (0 → sellable report)
| Sprint | Goal | Key metric |
|---|---|---|
| 0 | Setup & public data acquisition | Load a Sentinel-2 scene |
| 1 | First mineral-index maps | Visual correlation with known deposits |
| 2 | First ML classifier | AUC > 0.65 |
| 3 | Multi-sensor fusion | AUC > 0.72 |
| 4 | Vegetation as a sensor | +2% AUC from vegetation |
| 5 | Transfer learning | AUC > 0.60 on a new zone |
| 6 | Deposit-model integration | 3/5 known deposits found |
| 7 | First sellable report | Positive geologist feedback |
3. Recorded results — multi-zone, multi-deposit, multi-continent
Conservative, leakage-resistant spatial cross-validation AUC. Higher is better (0.5 = random).
| Zone | Deposit type | Baseline | Full fusion |
|---|---|---|---|
| Chuquicamata (Chile) | Porphyry Cu | 0.79 | 0.88 |
| Kalgoorlie (Australia) | Orogenic Au | 0.81 | 0.94 |
| Zambia Copperbelt | Sediment-hosted Cu | 0.74 | 0.76 |
| Peru | Porphyry Cu | 0.70 | 0.76 |
| Arizona (USA) | Porphyry Cu | 0.72 | 0.72 |
Where these figures come from. They are the best recorded configuration per zone, and the configurations differ between zones. The cross-domain comparison run — same 12 bands, no EMIT, no geophysics — records Chuquicamata 0.862, Kalgoorlie 0.769 and Zambia 0.615. An AUC is a ranking score inside one experiment: it is not a calibrated probability of mineralisation, and it carries no depth and no tonnage.
Fusion is not universal: it lifts strong-baseline zones and can hurt weak ones. The engine therefore selects features per zone and per deposit type — there is no single planet-wide formula.
4. Feature families — what each one contributed
| Family (public-access source) | Verdict |
|---|---|
| Satellite spectral baseline (Sentinel-2 / Landsat) | Core — always included |
| 20-year thermal climatology (Landsat) | Contributing input feature in these experiments — universal, modest lift. Not a validated detection method; the standalone thermal hypothesis is PARKED. |
| EMIT hyperspectral alteration (NASA) | Contributing input feature — porphyry Cu experiments only. The standalone spectral-residual hypothesis was NOT RUN. |
| Geology / lithology (Macrostrat) | Contributing input feature — selective, on three-zone evidence. |
| Aeromagnetics (national surveys) | Selective — small but real lift |
| Multi-decadal NDVI / vegetation (Landsat) | Selective — vegetated zones |
| Terrain / spatial gradients | Redundant with spectral — not added |
5. Key learnings
- Labels dominate. Same satellite data, Kalgoorlie went from 0.58 → 0.77 just by going from 16 to 205 curated deposit labels.
- Deposit type > commodity > geography. Training on pure types sharpens the model; mixing types confuses it.
- Fusion needs a floor. A zone needs a baseline AUC ≈ 0.73 before extra layers help rather than hurt.
- Transfer is zone-specific. Satellite features encode geography, not transferable geology — each zone earns its own stack.
- The gap is data, not ML. Surface screening is near its ceiling; the next gains come from new public-access layers.
6. Honest scorecard
We grade the engine against a fixed four-dimension objective. This is the real number, not a marketing one:
| Dimension | Score | Status |
|---|---|---|
| Mineral identity | 4.0 / 10 | Satellite encodes geography, not mineralogy |
| Depth | 4.1 / 10 | Needs subsurface geophysics (gravity / AEM) — access-blocked |
| Coordinates | 7.0 / 10 | ~30 m resolution, ~1 km² zones |
| Certainty | 7.7 / 10 | Calibrated Brier 0.1955 → 0.1711 (isotonic, n=211,592). Lower per-fold values exist in individual ablations; the calibrated model is the figure that is quoted. |
| Total | 22.8 / 40 | Internal rubric total: the sum of four grades from 0 to 10. It is not a percentage of potential, not a probability and not a measure of any deposit. |
Path forward. Each next step adds evidence a satellite cannot give: public geophysics, then drill-hole data, then a field campaign. A GeaSpirit report is designed to de-risk and prioritise that campaign before money is spent in the field.
7. Limitations
- Remote sensing and public data screen opportunity; only drilling confirms a deposit.
- Scores are prioritization indicators — not probabilities of economic mineralisation, and not valuations.
- Depth estimation is limited by access to subsurface geophysics, much of which sits behind manual portals.
- GeaSpirit provides intelligence and prioritization, never investment, legal or technical advice.
8. Parcel investigation, relief and resistivity (C2141)
- Any Spanish cadastral reference. GAIA resolves the official parcel, builds a study window and lists every public source covering it with an explicit state: loaded, available, download blocked (CAPTCHA), requires credentials, licence unclear, does not cover, not found, not applicable or failed. Two test parcels in Murcia were investigated end to end.
- Relief. The official 0.5 m terrain model (MDT, 3rd LiDAR coverage) exists for these parcels, but its download requires a CAPTCHA. GAIA lists the exact sheets; a person downloads them and GAIA uses them at their native cell. Until then the scene states that it uses the 5 m model. Zooming an image beyond its native resolution adds no detail, and the viewer says so.
- SAR + ERT (Buto study). We read Abouarab et al. 2026 (Acta Geophysica, doi 10.1007/s11600-026-01809-4): Sentinel-1 amplitude was used to choose where to measure; the buried structure was found by field ERT and confirmed by excavation. Their data and code are not published, so the result cannot be reproduced from outside. In GeaSpirit it remains NOT VALIDATED. No satellite replaces a field ERT survey.
- Our ERT chain. Synthetic solver test (24 electrodes at 3 m, buried block 3–6 m deep): position and depth recovered within 0.4 m, but the smooth inversion underestimates the resistivity contrast about threefold. On real field data with an independent borehole (C2135) the contrast was found but the contact depth was off by 11.5 m: failed by the preregistered criterion. A synthetic pass is not field evidence.

Two published techniques for hidden structures: Giza (2025) and Buto (2026)
Giza (Schumacher et al., Scientific Reports, 2025, open access CC BY 4.0). To confirm the ScanPyramids North Face Corridor behind the Chevron blocks of Khufu's Pyramid, the team measured the stone with three contact instruments: ground penetrating radar (antennas 200–800 MHz, reflections of electromagnetic waves), ultrasonic testing (25 kHz shear-wave pulse-echo array, reflections of elastic waves) and electrical resistivity tomography (stainless-steel sheet electrodes on wet sponges, dipole–dipole on ten parallel lines 0.25 m apart, inverted in 3D with the open-source pyGIMLi). Each method gives a 2D image of the first 2 m. The images were registered, cropped, resampled to one pixel size and normalised to 0–1, then merged with a discrete wavelet transform (wavelet sym3, level 7; minimum of the approximation coefficients, maximum of the detail coefficients). The settings were chosen by visual inspection and the fused images were interpreted by a team of experts; the corridor had already been seen with an endoscope in 2023. The data are available only on request.
Buto (Abouarab et al., Acta Geophysica, 2026). Sentinel-1 radar images from orbit were used to choose where to measure in the Nile Delta, and ERT on the ground, inverted in 3D, located a buried Saite-period structure later checked by excavation. Data and code are not published.
| Technique | Giza 2025 | Buto 2026 | What GAIA can do with it |
|---|---|---|---|
| ERT — electrical resistivity tomography | Yes | Yes | Process existing field data (2D inversion tested on synthetic data only); needs electrodes on the ground |
| GPR — ground penetrating radar | Yes | Not the main method | No processing chain; needs an instrument on the surface |
| Ultrasonic testing | Yes | No | Not applicable to remote work; contact instrument on rock or structures |
| Sentinel-1 satellite radar (SAR) | No | Yes | Remote: finds the Sentinel-1 scenes over any parcel; used to choose where to look, never as a subsurface measurement |
| Wavelet image fusion (DWT) | Yes | No | Implemented as the paper describes, for co-registered layers of the same place; a visual aid, EXPERIMENTAL |
| 3D inversion | Yes | Yes | Not demonstrated in GeaSpirit |
What this means for remote work. Of the four measuring techniques only satellite radar is remote. Radar from orbit sees the surface: roughness, moisture and millimetre ground motion, not the inside of the ground. GPR, ultrasound and ERT need an instrument on site; GAIA can only process campaigns that already exist. The fusion step, on the other hand, is a mathematical tool that can combine any images covering the same place, including remote layers (satellite radar, spectral indices, relief, magnetic or gravity grids). GAIA refuses to fuse layers that do not cover the same window or that would have to be enlarged beyond their native resolution.
Our own test (preregistered, synthetic). On 50 synthetic scenes the published rule (sym3, level 7, min/max) produced an image better than the average of its two inputs in 50 of 50 scenes, but worse than the best single input (median AUC 0.86 against 0.97). Result: FAIL against the preregistered criterion. The fusion helps to look at two images at once; it does not detect anything by itself. Both techniques remain NOT VALIDATED in GeaSpirit.
Schumacher T. et al. (2025) Sci. Rep. 15, doi 10.1038/s41598-025-91115-8 (CC BY 4.0) · Abouarab et al. (2026) Acta Geophysica, doi 10.1007/s11600-026-01809-4
8. Where this is going
- Screening reports — where a zone ranks, on what evidence, with what is missing. A ranking is not a statement that a mineral is present.
- Global coverage — scale the public-access database toward every mine and prospect on Earth.
- Mining-asset marketplace — connect owners, explorers and investors (individuals & institutions) around evidence-scored assets and potential lands, with verifiable, timestamped provenance.