Search less ground. Find more deposits. On every continent.

Share of proven deposits in the best 10% of ground; random search finds 10%. Green: 95% interval above 10%. G9 ran on 88,815 records; G62 re-ran the same ranking on the 238,872 records of its date, with more deposits, mostly from national databases. Models were trained without the continent they rank (leave-one-continent-out). The US row of the model columns is empty because US models were tested separately on hidden US regions (below).

Lift = share of hidden deposits in the top 10% of ground ÷ 10%. Green: the 95% interval stays above random. Red: not distinguishable from random.

Pooled primary groups (Cr, Cu, Mn), all regions: 19.6% of 954 deposits in the top 10% (2.0×). Resampling whole regions instead of single deposits, which accounts for deposits clustering together, gives 12.8–26.3% overall and 15.8–35.2% for the US. Tin is not shown: 76 of its 101 hidden deposits sit in one Bolivian region, so it says nothing general.

Three quarters of the world's openly documented, well-located deposits are in US records. A model trained on them learns what US mineral belts look like in a coarse (4 km) magnetic map, and that does not carry over to the Andes or Central Asia. On the locked test outside the US, the ranking found 11% of 421 hidden deposits in the top 10%, which is what chance gives.

The leave-one-continent-out test confirms it: models trained without a continent ranked it at chance overall (10% with magnetics + gravity). For Andean copper they did worse than random, because porphyry belts there look different in coarse geophysics from those in the US.

What this means for a project outside the US:

• Known-showing ranking first. Validated on every continent.
• The geophysics quick scan is labelled experimental there. Every alarm shows the tested skill for its region.
• The value comes from the deep dive. Satellite mineral mapping (Sentinel-2, ASTER, EMIT/EnMAP hyperspectral) at the scale of your licence, compared with background in the same scene.
• Field checks by local geologists, with targets locked before anyone goes out, so hits and misses both count.
• Regional training data from national surveys as it becomes available. Each new region is tested the same way before we claim anything.

Mountain Pass, Mount Weld, Kızılçaören and Bayan Obo were removed from all training. In a 20 × 20 km window around each, neither the magnetics model nor the known-showing ranking put the deposit in the top 5% of cells. Mountain Pass came closest (top 10% with magnetics). For Kızılçaören and Bayan Obo no other public rare-earth showing lies within 50 km, so the showing ranking gives no answer there by design.

At licence scale the evidence has to come from satellite mineral mapping and field checks. Files: analysis/g4_cases/.

On 13 February 2024 the heap-leach pad at the Çöpler mine (İliç, Türkiye) failed. Using only Sentinel-1 radar from before the failure and an alert rule written down in advance, the rule first fired on 2023-09-03, 163 days earlier.

We then ran the same rule on 161 other mines on four continents. It fired at 9% of them overall, but at 5 of 8 heap-leach mines without a failure. Kışladağ, a heap-leach mine the size of Çöpler, alerted 17 times.

We then tested a second rule that compares each pad with its own history. It stopped the heap-leach false alarms (1 of 8) but never fired at Çöpler: the pad moved steadily, without visible acceleration, in 12-day radar pairs.

So radar here measures the normal settling of leach pads, which is useful monitoring, but it is not a failure warning. Warning needs denser data (ground instruments, daily commercial radar) and more failure cases.

The ranking is only as good as the public records around a licence. Adding one free national database and re-testing on the same deposits:

• Australia (Geoscience Australia OZMIN): 43.4% → 63.9% of proven deposits in the best 10% of ground (gain +20.5 points, +10.3 to +32.7).
• British Columbia (MINFILE): 11.0% → 47.3% (gain +36.3 points, +23.4 to +46.5).
• Brazil (Serviço Geológico do Brasil): 24.3% → 61.0% (gain +36.7 points, +22.9 to +48.3).
• Sweden (SGU Mineralresurser, tested on independent FODD deposits): 4.7% → 62.6% (gain +58.0 points, +36.9 to +69.0).

So in every new country the first step is its national occurrence database. The product already includes these four.

On Australian ground with no known showing within 50 km, a model on airborne geophysics and the national geology map put 56.2% of gold and copper deposit cells into the best 10% (G36). Adding state-recorded sites as training labels raised it to 66.0% (+9.8 points, +0.6 to +17.6; G42), and treating national parks as unknown rather than barren to 68.5% (G45). Random: 10%.

Independent check. Candidate cells chosen where our catalogue showed nothing were compared with random cells from the same ground, against state databases the model never used: Western Australia 16.1× (8.3–37.7; G40); Northern Territory, Queensland and NSW 11.5× (5.0–34.0; G41). South Australia, Victoria and Tasmania: 3.4× with an interval including no effect (G43, not passed).

Two challenges to that result (G55). A simple expert rule (greenstone or volcanic host rocks near a mapped fault) reaches 45.9%; the model beats it by +22.7 points (+11.8 to +32.1). A model trained on labels shuffled within 1-degree blocks, which knows only which areas are fertile, still reaches 61.9%; the model's lead over it is +6.6 points (+1.4 to +12.7). So most of the skill is regional, and the cell-level part is real but modest.

How sure is each candidate (G58). On the same ground, cells in the model's top 1% held a record of the metal the model had not seen 12.0× as often as the average cell (9.5–15.1); the 1–5% band 7.1× and the 5–10% band 3.5×, so a top-1% cell is 3.4× as likely to hold one as a 5–10% cell (2.6–4.5). Retraining the model 25 times on resampled ground gives each cell a spread of ranks: at the same rank, cells the models agreed on held a record 1.32× as often as cells they disagreed on (1.13–1.50; both passed). So a candidate's band and the models' agreement are both shown as its confidence. Two limits: with catalogue deposits alone (far fewer) the agreement effect is not significant (1.10, 0.74–1.83), and the figures differ by state (top 10%: Western Australia 4.2×, Queensland 5.7×), so they are given per state.

Québec, predicting later discoveries (G54). Trained only on what was on record before 2005, a model on the national airborne magnetics, gravity and the provincial geology map put 51.8% of the gold and copper finds made since 2005 on ground at least 10 km from any earlier record into the best 10% of that ground (37.8%–63.6%; passed). Two registered checks limit the claim: simply ranking by distance to earlier finds reached 40.3% (the model's lead, +11.5 points, is not significant), and a model trained on labels shuffled within 1-degree blocks still reached 45.3%. So most of the skill is knowing which belts are fertile; within a belt, the evidence is weak.

Brazil, without airborne data (G57). Trained on SGB records registered before 2008, a model on the national 1:1M geology map, faults and global grids put 55.3% of later gold and copper records on new ground into the best 10% (43.9%–65.9%; passed). Unlike Québec, it clearly beats block-shuffled labels (+16.4 points), but simply looking next to earlier records reached 62.3%, and the model does not beat that.

Inside a licence, with the licence holder's own data (G59). In 115 Western Australian project areas (open-file company data: 2,592 cells later drilled by RC or diamond holes), first-pass geochemistry reported before that drilling (shallow holes and surface samples) ranked the 500 m cells where the later holes hit: the explorer's simple rule, drill the highest first-pass value, did better than chance (AUC 0.58, 0.54–0.63; random 0.50). A model trained on other projects did not beat that rule (-0.02 AUC, -0.07 to +0.03). So inside a licence we organise, check and map the holder's data and keep the simple rule; we do not claim a better ranking there.

Africa, public continental layers (G60). Inside Africa, the USGS continental geology map, 90 m surficial lithology and global grids put 56.4% of gold and copper deposit cells of held-out 2-degree blocks into the best 10% (48.7%–63.2%; passed). Most of it is regional: labels shuffled within blocks reach 48.6% (the model leads by +7.8 points, +2.2 to +13.5), and where showings exist, distance to them ranks better (76.5%). Highest in the Copperbelt (copper 80.0%) and West African gold (64.7%), but only Northern Africa beats its regional control clearly, so that is the one candidate list the registration allows.

Australia at 1 km instead of 11 km (G64). Splitting each 0.1-degree cell into 1 km cells and ranking them with fine airborne magnetics, radiometrics, gravity, lithology and terrain put 33.4% of held-out gold and copper deposit cells into the best 1% of the area, against 27.3% for the 0.1-degree model (+6.1 points, +2.8 to +9.3; passed). Half the deposits fit in 39,085 km² instead of 48,773 km². The gain holds away from mapped mines (+4.0 points) and on ground with no showing within 50 km (+8.7 points), so it is not mine recognition. Gold and iron carry most of it; copper barely moves.

Ground drilled without result counts too (G65). In Western Australia, adding a cell's exploration history (earlier drilling and sampling, including holes that found nothing) to the 1 km model ranked the cells later drilling hit better: AUC 0.669 to 0.713 (+0.044, +0.028 to +0.063; passed). Cells drilled barren later hit at 3.9%, against 6.4% elsewhere: less often, not never. A flag for cells unlike the training data did not mark weaker ranking (failed), so we do not show it as a warning.

On data three geological surveys prepared (G67). We ran our protocol on the Geological Survey of Canada, USGS and Geoscience Australia zinc-lead datacube (5.16 million cells), holding out whole 5-degree regions. Adding distance to known occurrences to the published-style geology and geophysics model raised capture in the best 6% of area from 37% to 54% (+18 points, +8 to +29; passed). Distance to occurrences alone ranked best (AUC 0.92 for Mississippi Valley-type, 0.97 for clastic-dominated). The published gradient-boosting accuracy (0.98–0.99, random test split) did not hold under held-out regions (0.64 and 0.89); weights of evidence matched its published values (0.80 and 0.86). Same answer as our own tests: known showings first, models where there are none.

What did not add (best on unexplored ground stays 68.5%; random = 10%):

Candidate lists (gold, copper, iron, nickel-cobalt) are locked by SHA-256 in the prediction register before anyone checks them; a monthly forward check counts new state records in the candidates and in control cells fixed at the same time. Cells are about 11 km: they say which part of a region to look at, not where to drill. In Australia, 1 km cells rank better (G64), which narrows the district, still not the hole.

In Australia, a separate model for each commodity on known showings, airborne magnetics and radiometrics put 61.5% of deposit cells into the best 10% of the ground, against 49.7% for showings alone (gain +11.8 points, +6.3 to +17.3; G25, reproduced). Adding the national 1:1M geology map (rock types and faults) raised it to 68.1% (+6.6 points, +4.2 to +9.2; G29).

A pre-registered tournament of eight methods with a multiple-comparison correction (G26): positive-unlabelled learning was best (63.2%), then random forest and XGBoost; a neural network, weights of evidence and analog search did not beat showings.

Largest gains for iron, gold and copper; rare earths improve only once rock types are added. Areas of about 11 km, not targets inside a licence.

On 814 producing deposits, Sentinel-2 and terrain features told the deposit from its own surroundings, but mostly by recognising the mine itself (0.709 where a mapped mine lies within 1 km, 0.568 away from one; 0.5 = chance; G27). On 1,229 undeveloped deposits and prospects with no mine within 5 km the signal is weak but real: 0.521 (corrected interval 0.502–0.557), stronger away from mines and for base metals (G28, reproduced).

Used as a supporting layer inside a licence for base metals, never as a target on its own.

Geoscience Australia's free 80 m magnetic grid against the global 4 km grids, with the same model and block cross-validation, on 2,956 cells holding Australian mines and deposits. Share in the best 10% of ground: global grids 23.4%, airborne 30.7%, both 32.5% (gain +9.1 points, +4.8 to +13.5). A re-run from scratch gave identical numbers.

Largest gains for iron, nickel-cobalt, copper and gold; none for rare earths.

On top of known showings (G16): showings alone 45.5%; with airborne magnetics 48.8% (gain +3.3 points, +0.5 to +6.3). A small, real addition; showings do most of the work, and for rare earths showings alone are better.

Using only the showings recorded before 1960, the ranking put 17.7% of 875 deposits recorded in 1960 or later into the best 10% of the ground (13.2%–22.1%), about 1.8× random. Two thirds of them are outside the US.

Only about 7% of showings carry a date, so this understates what the full catalogue gives today (35.5%, G9).

Trained on other continents, our geophysics models rank a new continent at chance. Trained on blocks of the same continent, they ranked its unseen blocks at 23% in the top 10% of ground (17%–30%), and a second, independent set of blocks repeated it at 38% (16%–51%). Random search finds 10%.

Where public showings are sparse (manganese in Latin America) the models beat the showing ranking clearly. Where showings are dense, adding the model to them did not help (42% vs 38%). A rule that switches by showing density gained +10 points but did not pass (-7 to +25).

Caveat: part of the second set is one Bolivian tin district. Every new region still needs its own training data and its own test.

We mapped alteration minerals (alunite, kaolinite, muscovite, jarosite) around 998 undeveloped deposits outside the US in 279 EMIT scenes and compared each with its own ground 3–10 km away. Deposits came out ahead in 51.0% of comparisons (49.2%–52.7%): chance is 50%.

A second test on 144 new scenes checked the one class that had looked promising (carbonate): 52.1% (49.2%–55.0%), chance again.

The EMIT layer stays in the product as a mineral map for geologists, not as ranked targets.

Our rare-earth machine-learning map failed its blind test (G1). A hint that it worked on alkaline-igneous deposits (0.54 on 8 sites) did not repeat on 30 fresh ones: they beat random points in 48.8% of comparisons (41.0%–57.2%), chance again.

Supply Watch, the activity index for producing mines, did not track Mountain Pass production 2019–2024 (Spearman -0.20). It stays labelled beta.

• It does not show that a high-ranked cell holds an economic deposit. Rankings are places to check, not resources or reserves.
• It does not pick drill holes. Among 18,600 already-drilled holes in Western Australia, neither ranking separated rich holes from barren ones (G5, failed): the rankings choose ground at the 1–10 km scale, not holes inside a prospect. A later test on 41,220 follow-up holes, with first-pass values near each collar and a model, gave the same answer (AUC 0.51–0.55; G66).
• It is not yet proven in Türkiye on public data alone: on 64 known Turkish deposits it reached 2.2× random, but the interval includes chance, because three quarters have no public showing nearby (G18). National data (MTA) is the fix, as in four other countries. With the USGS porphyry assessments added (G61), the showing ranking clears chance in Türkiye (39.4%, 18.5%–59.2%), mostly on porphyry deposits ranked by prospects from the same assessments; ground without showings is still at chance there.
• It does not yet improve rare-earth ranking with geophysics. Adding Australia's airborne thorium grid to showings lowered the share of REE cells in the best 10% from 48.3% to 32.2% (G21, failed, 118 cells); thorium alone reached 25.4%, above the 10% of chance, but not enough to add on top of showings.
• It does not find "look-alike" ground by simple similarity. Ranking cells by how close their airborne magnetic and thorium signature is to known deposits put only 6.0% of deposit cells in the best 10%, below chance (G24, failed): deposits are unusual, and plain similarity favours ordinary ground.
• Newer layers do not always add: derived geology and geophysics with positive-unlabelled learning reached 70.0% but did not significantly beat the geology-map model (G30); national sediment geochemistry added nothing on top of geology (G33, -0.5 points).
• A model trained in Australia did not beat British Columbia's own showings (27.6% vs 32.8%, gold; G31). Without any showings it still reached 31.0%, a lead for ground with no public records.
• Known-showing ranking helps most in explored districts. Where no public showing lies within 50 km, it says nothing, and we tell you so.
• "Deposit" here means a documented producer or defined resource in public records; undiscovered deposits cannot be counted.
• The test regions were squares that hold known deposits: mineral districts. Ranking inside a district is harder than telling a district from a desert, and is the decision explorers actually face.
• The test split is now used for these model versions. Future models are judged on a separate holdout and new data.

• Global tests: analysis/global_test/prereg_g9.json (known showings), prereg.json (leave-one-continent-out); scripts 17_showing_prior_global.py, 15_global_transfer_test.py
• Baselines: analysis/16_g2_baselines.py, results discovery_test/g2_baselines*.csv
• US blind test registration: analysis/discovery_test/prereg.json, model fingerprints model_hashes.txt
• Script: analysis/13_discovery_test.py; results results.json, per_group.csv, per-cell scores cells_<group>.csv
• Data: dataset ds-2026.10 (46,765 positives from USGS MRDS and national sources, whole-block splits), EMAG2v3 magnetics (NOAA), Natural Earth land
• Every test from G36 on: analysis/g<NN>_test/prereg_g<NN>.json (written before the data were opened) and results_g<NN>.json; all 70 with their verdicts in the test register
• Data behind the tests: about 74 GB of public data, mostly national airborne grids for Australia (about 43 GB), continental layers for Africa (about 15 GB), Western Australian open-file company reports (6 GB) and the three surveys' zinc-lead datacube (2.5 GB). Power comes from the number of independent regions held out, not the gigabytes. Third-party data is not redistributed where its licence does not allow it; only results are published.
• InSAR: analysis/insar/copler/prereg.json, prereg_controls.json, ASF HyP3 products

Independent review is welcome: we share the full package with academic partners on request.