All data → one dataset → analysis → forecast → validation targets
How Cratona turns a territory's scattered geodata into a verifiable forecast and a short list of targets.
We compare seven architectures — from Random Forest to convolutional networks with FiLM geochemistry fusion — and work as an ensemble. Convolutional models see the texture and arrangement of features, not just their averages, and transfer better to new territories.
| Model | Architecture | Input | Geochemistry | Uncertainty |
|---|---|---|---|---|
| M0a | Random Forest (500 trees) | 108 = 54 ch × (mean + std) | none | tree std |
| M0b | Random Forest (500 trees) | 128 = 108 + 15 geochem | concatenation | tree std |
| M1a | FCNN 108→128→64→1 | 108 | none | MC-Dropout ×20 |
| M1b | FCNN 128→128→64→1 | 108 + 15 geochem | concatenation | MC-Dropout ×20 |
| M2a | CNN, 54 channels | 64×64×54 | none | MC-Dropout ×20 |
| M2b | CNN, 54 channels | 64×64×54 | concatenation | MC-Dropout ×20 |
| M2c | CNN + FiLM | 64×64×54 | FiLM | MC-Dropout ×20 |
The model runs the prediction 20 times with a random part of the network switched off. If the results repeat, the forecast is considered stable. If the estimates differ noticeably, the area is flagged as a zone of elevated uncertainty.
The model is trained on some regions and tested on others that were not used in training. This shows how robustly it performs beyond the source data and how well the learned patterns transfer to new areas.
The ConvNeXt architecture with geochemical features integrated via the FiLM mechanism showed the best separation of ore-bearing and background areas among the tested models. The score distributions of the two classes overlap less, so the model ranks the territory by prospectivity more robustly. Confidence is assessed separately, from prediction stability and the uncertainty level.
AUC-ROC shows how well the model ranks the territory by prospectivity. The higher it is, the more often areas with known mineralisation are rated above background ground. Model quality is measured on territories that were never used in training.
