AI USE CASE
Mineral Grade Prediction from Drill Data
Predict ore mineral grades in real-time using deep learning on drill core imagery and assay data.
What it is
This use case applies deep learning and predictive analytics to drill core images and assay results to estimate mineral grades across the ore body continuously. By replacing or augmenting slow laboratory workflows, mines can reduce grade estimation turnaround by 50–70%, enabling faster blast-hole decisions and reducing ore dilution. Improved grade control typically delivers 3–8% uplift in mill feed quality, translating directly to higher recovery rates and lower processing costs. The system integrates with existing drill-and-blast planning tools to feed updated grade models in near real-time.
Data you need
Historical drill core imagery (preferably hyperspectral or high-resolution RGB), matched assay results, and georeferenced drill hole collar data across multiple ore zones.
Required systems
- data warehouse
- erp
Why it works
- Engage senior geologists early to co-design the labelling and validation workflow, building trust in model outputs.
- Standardise core photography protocols (lighting, resolution, depth tagging) before training data collection begins.
- Start with a single well-characterised ore zone to demonstrate predictive accuracy before broader rollout.
- Integrate grade predictions directly into the blast-hole planning software so value is visible within normal workflows.
How this goes wrong
- Insufficient or inconsistent historical assay data makes training unreliable, producing poorly calibrated grade predictions.
- Core imagery quality varies across drilling crews or equipment, introducing noise that degrades model performance.
- Geologists distrust model outputs and revert to manual estimation, preventing operational integration.
- Model trained on one ore zone fails to generalise when applied to geologically distinct zones within the same deposit.
When NOT to do this
Do not deploy this system if your mine lacks at least two years of spatially registered assay records paired with core imagery — the model will be too data-sparse to outperform simple kriging interpolation.
Vendors to consider
Sources
This use case is part of a larger Data & AI catalog built from 50+ enterprise transformation programs. Take the free diagnostic to see how it ranks against your specific context.