Introduction
Geologists have always collected more data than they could analyse. In Ethiopia’s Adola Greenstone Belt, burlap sacks of soil samples dating back to the 1980s sit in storage sheds – mute witnesses to half‑forgotten exploration campaigns. Enter machine learning (ML). By feeding those legacy assays into gradient‑boosting models, today’s explorers convert “dead” data into probabilistic heat‑maps that show where to drill next. Adola Goldfields, the Moyale‑based venture majority‑owned by Sowat Ltd., is spearheading the approach in East Africa.
What’s in a soil sample?
A standard minus‑80‑mesh soil sample captures trace‑element halos – gold, arsenic, antimony, bismuth – above concealed ore bodies. Historical campaigns analysed fewer pathfinders and stored the chemistry in paper ledgers. Modern multi‑element ICP‑MS analyses expand the vector library from four elements to forty‑eight, but the real breakthrough comes when ML algorithms detect non‑linear relationships humans miss.
The AI pipeline demystified
| Stage | Legacy practice | AI‑enhanced workflow | Value add |
|---|---|---|---|
| Data rescue | Manual re‑entry of assay ledgers | OCR + regex autoload into cloud DB | 90 % time‑saving |
| QA/QC | Spot‑check duplicates | ML outlier detection & drift analysis | Detects hidden lab bias |
| Feature engineering | Au, As thresholds | 48‑element ratios, PCA, spatial lags | 6 × more signal |
| Modelling | Kriging or simple ratio maps | Gradient‑boost or CNN on raster stacks | 3–4 × higher AUC |
| Target ranking | Geologist intuition | Probability heat‑maps + economic filters | 60 % fewer false positives |
| Drill‑pad design | Straight‑line fences | Algorithmic optimal‑spoke pattern | 35 % fewer metres drilled |
Sources: GoldSpot case study Investing News Network (INN); geospatial AI research ScienceDirect.
Algorithms in the wild
Extra‑Trees & XGBoost – KoBold Metals uses ensemble decision trees to weigh geophysical, geochemical and structural predictors WIRED.
Convolutional Neural Nets (CNN) – Researchers in Geology (Mar 2025) coupled mineral‑system knowledge with CNNs to hit 0.89 AUC in gold prospectivity Geoscience World.
Graph Neural Nets (GNN) – Experimental work links sample points as nodes in a geological graph, improving small‑data performance.
Proof‑points from industry
GoldSpot Discoveries / New Found Gold – AI targeting helped drill 19 m @ 92.9 g/t Au in Newfoundland Investing News Network (INN).
Rio Tinto – Centralised 100‑years of exploration data in Seequent’s cloud, cutting data retrieval time from weeks to hours Seequent.
KoBold Metals – Paid $15 m for historic Alaskan nickel‑copper data, believing ML would unlock blind deposits Alaska Energy MetalsAxios.
Adola Goldfields’ implementation
Digitisation – 12 000 legacy Moyale soil samples scanned; OCR achieved 98 % accuracy post‑validation.
Model training – 48‑element dataset plus Sentinel‑2 bands fed into CatBoost; cross‑validated AUC 0.83.
Output – Priority corridor 3.2 km × 900 m with 26 % probability > 0.5 g/t Au down‑hole.
Next steps – 3 500 m RC drill programme scheduled Q4 2025; success will feed valuation ahead of Canadian RTO.
Exploration VP Henry Potgieter calls the pipeline “a $2‑per‑metre brain that never sleeps”.
Economic impact
A typical greenfield campaign spends $300 per drilled metre. Raising the hit rate from 10 % to 25 % cuts the discovery‑phase metre count by 60 %. For Adola’s initial 20 koz oxide target, that’s $1.8 m saved – equal to 9 % of the current $20 m capex budget.
References:
• ScienceDirect – Integrating Soil Geochemistry & Machine Learning (2024)
Peer‑reviewed study demonstrating ML uplift in gold deposit prediction.
• KoBold Metals – Company Overview (accessed 2025)
Describes AI‑first approach to battery‑metal exploration.
• WIRED – Algorithms Hunting for Battery‑Metal Mother Lode (2023)
Explains KoBold’s ML workflow on historical datasets.
• Axios – KoBold & BHP Use AI for Exploration (2021)
Early media on AI shaping joint‑venture targeting.
• ScienceDirect – Geospatial AI Prospectivity Mapping (2024)
Describes geographically weighted logistic regression with neural nets.
• Seequent – Rio Tinto Tackles Exploration Data Management (2024)
Highlights value of centralising historic drill & soil datasets.
• Alaska Energy Metals – Sale of Exploration Data to KoBold (2023)
Shows real $$$ value assigned to legacy data libraries.
