AI-Driven Continental-Scale Mineral Prospectivity Engine
Commercial licensing available. Research use permitted. See LICENSE for terms.
GeoAnomalyMapper uses a Physics-Informed Neural Network (PINN) to solve the gravity inverse problem at continental scale — converting Bouguer gravity anomaly data into subsurface density contrast models, then extracting ranked prospectivity targets.
Bouguer Gravity → Residual Separation → PINN Inversion → Dual-Pipeline Extraction → Scored Targets
The system runs a dual pipeline — finding both mass-excess anomalies (VMS, IOCG, skarns, Ni-Cu) and mass-deficit anomalies (epithermal gold, alteration halos, kimberlites) — validated against independent geochemical and MRDS datasets.
| Metric | Result |
|---|---|
| Targets generated (continental US) | 1,634 |
| Tier 1 high-confidence targets | 31 |
| Geochemical enrichment vs. baseline | 8.5x (32.4% hit rate vs. 3.8% random) |
| Negative control specificity | 100% (zero false positives in barren regions) |
| Statistical significance | >7 sigma |
Independent validation used the NURE geochemical database (397,000+ sediment samples) — completely separate from training data. See docs/SCIENTIFIC_VALIDATION_REPORT.md for full methodology.
Regional Screening Packs — ranked prospectivity leads + evidence bundle per commodity/district. Suitable for early-stage teams selecting which districts to enter.
"Bring Your Own Data" Inversion Service — you supply higher-resolution gravity/mag, we deliver ranked anomalies, uncertainty maps, and an HTML+GeoJSON target package. Suitable for juniors with geophysics data but no internal AI inversion pipeline.
Automated Due Diligence Reports — submit coordinates, receive a go/no-go desk report: MRDS cross-check, geochemical coverage flags, density contrast evidence. Suitable for exploration managers screening incoming deal flow.
Commercial Software License — deploy the pipeline on your own infrastructure.
For pricing and availability, open a GitHub issue or contact via the profile.
Core model: DensityUNet with physics layer implementing Parker-Oldenburg forward gravity. Loss function: data fidelity + structural coupling (EMAG2 magnetic regularization) + sparsity.
Resolution: ~2km continental grid. Optimized for district-scale targets (1–20km). Not suitable for resolving individual deposits <500m. Resolution limit is a documented design constraint, not a bug — the filter is calibrated to district scale by design.
Depth ambiguity: All gravity inversions are non-unique. Outputs are prospectivity indicators, not structural models. Ground truthing is required.
Deposit types detected:
- Mass-excess: IOCG, VMS, skarns, magmatic Ni-Cu, dense intrusives
- Mass-deficit: Epithermal gold systems, alteration halos, sediment-hosted Au, kimberlite pipes
GeoAnomalyMapper/
├── pinn_gravity_inversion.py # PINN architecture (DensityUNet + physics layer)
├── loss_functions.py # Custom loss (structure-guided TV + magnetic coupling)
├── train_usa_pinn.py # Training pipeline
├── predict_usa.py # Continental-scale sliding window inference
├── extract_dual_targets.py # Dual-pipeline target extraction
├── phase2_validation.py # MRDS cross-reference + confidence scoring
├── verify_skeptic_v2.py # Forensic validation + negative controls
├── data/outputs/ # Target CSVs, scored lists
└── docs/ # Scientific validation, forensic audit, methodology
pip install -r requirements.txt
python train_usa_pinn.py
python predict_usa.py
python extract_dual_targets.py data/outputs/usa_density_model.tif
python phase2_validation.pyGravity input data: USGS Bouguer anomaly grid. Magnetic data: EMAG2.
Both are publicly available. See setup_usgs_data.py for download helpers.
- Continental public grids (~2km) cannot resolve deposits smaller than ~500m.
- Basin & Range regional gravity lows (Nevada) mask discrete epithermal anomalies at this resolution.
- Gravity inversion is inherently non-unique (depth vs. density ambiguity).
- Outputs are statistical prospectivity indicators — not drill targets. Field verification required.
The validation suite (negative controls, geochemical enrichment, forensic audit) documents
exactly where the model works and where it doesn't. See docs/ for the full record.
Proprietary. Commercial use requires a license agreement. Research use permitted. See LICENSE for terms.
- USGS for public gravity and MRDS data
- NURE program for geochemical validation data
- The geophysics and mineral exploration research community