Overview
In this step, you’ll select a Data Fusion model and set the number of training epochs to embed your input features. This process reduces the complexity of your data while tailoring the model to your specific geological setting.
Why do we Embed Input Features?
Embedding input features means transforming your selected rasters (up to 64) into a set of 24 lower-dimensional representations (embeddings) that preserve meaningful spatial patterns and relationships. This is done using a Data Fusion Model, which is a pre-trained AI model trained on hundreds of thousands of rasters across specific mineral system types.
The embedding process captures:
Spatial relationships across rasters (how different layers relate to one another)
Spatial structure within each raster (how features are arranged in space)
The Data Fusion Model, trained on large-scale data, is then fine-tuned with your selected rasters and learning points. This produces 24 new embedded rasters that are optimized for your AOI and geological context.
Why This Step Matters
Geoscience Perspective
Embedding helps isolate geological patterns that might not be immediately visible in raw data. By using a Data Fusion model trained on similar mineral systems, you can leverage global insights and tailor them to your specific project area. This increases confidence in your predictions and helps ensure the model is grounded in domain-relevant geoscience.
View the full list of available Data Fusion models and the mineral systems they are trained on here.
AI Perspective
This step applies dimensionality reduction to your dataset, making it more manageable for machine learning while preserving essential spatial and structural features.
Selecting the right combination of rasters improves the model’s ability to recognize geological patterns during training. The number of epochs you set controls how much the model learns from your local data, allowing it to fine-tune predictions specific to your project. DORA includes an early stopping mechanism that automatically halts training once performance stops improving, helping to avoid overfitting and unnecessary computation.
This is a form of few-shot learning. The model has already learned from hundreds of thousands of rasters across global projects, but it adapts to your selected input features by learning just enough to optimize predictions for your geological setting.
💡Tip: When adjusting the number of training epochs, consider the trade-offs:
|
Step by Step Instructions
Open Embed Input Features
From the experiment setup panel, click Step 4: Embed Input Features.
Select a Data Fusion Model
From the dropdown, choose the model that aligns with your target mineral system.
Not sure? Start with Master_Model.
Select None if your deposit is highly unique and you want to train embeddings from scratch.
Set the Number of Training Epochs
Enter the number of training epochs. This determines how much the model will fine-tune itself on your selected data.
The default is 100.
Training may stop early if the model’s performance stops improving.
Click Generate Embeddings
This triggers the dimensionality reduction and creates rasterized outputs for each input feature.
(Optional) Review the Visual Results
Use the 3D Layers List to preview the generated raster layers. Toggle visibility (eye icon) and adjust the color scheme or base displacement scale to better visualize textures and surface patterns.
While reviewing, look for signs of artifacts — visual inconsistencies caused by mixing high-resolution datasets with coarser grids that cover the full AOI. These can appear as abrupt edges or linear patterns within the embedded layers.
These artifacts may lead to skewed VRIFY Prospectivity Score (VPS) results, such as overly prospective areas around the edges of smaller surveys.
If this happens, try removing the problematic layer and rerun the embeddings to see if the issue resolves.
Complete Step
Click Proceed to move to the next stage.
Tips & Considerations
Data Fusion Model Naming
Model names like c2p4 refer to the number of training images used (e.g., c2p4 represents 24,000 training images).
Prioritize models trained on the correct deposit type, then select the one with the highest training count for the most reliable results.
Did you notice the naming convention c2p4, c1p8, etc? It’s a reference to C-3PO from Star Wars!
Epoch Strategy
If results aren’t improving with additional epochs, stick with the early-stopped model — it’s likely already optimized.
If you select the None Data Fusion Model, consider using a higher number of epochs (e.g., 300) to allow for deeper refinement, since there are fewer pre-trained images available to guide the embedding process.
Learn More
Previous Steps:
Next Steps:
Still Have Questions?
Reach out to your dedicated DORA contact or email support@VRIFY.com for more information.