Minerals Processing Optimization
What the AI does
AI maps and categorizes ore types, generates synthetic samples to model chemical/mineralogical properties without physical testing, and simulates the grinding + flotation chain to recommend optimal additive doses. Result: higher yield, less energy, fewer hazardous chemicals.
Domain
Mining and minerals processing (copper, nickel, lithium, rare earths).
Pattern
- In: assay data + ore composition + historical processing results + sensor data from the plant.
- Reasoning: characterization model categorizes incoming ore; simulator models chemical separation under candidate parameter sets (synthetic-data-generation fills the gap where physical testing is too slow/destructive).
- Out: real-time setpoint recommendations for the processing plant; pre-emptive flagging of ores that need parameter adjustment.
- Human checkpoint: plant operator confirms parameter changes; metallurgist reviews on novel ore arrivals.
Deployments in the wild
- Generic ER&I pattern in deloitte-ai-dossier-eri; no single named deployment.
What’s actually being measured
- Yield improvement (asserted; no quantified figure).
- Reduction in energy and reagent consumption (asserted).
- Not measured publicly: failure rate on novel ore types — Deloitte itself flags “novel ores” as the canonical OOD-failure example.
Concept linkages
- scaling-wall — the textbook in-source example: “models struggle to generalize when faced with novel ores.”
- synthetic-data-generation — the dark-data workaround that produces this OOD risk.
- agentic-revolution — moving from suggested-setpoint to closed-loop control is the agentic step.
- autonomy-slider — closed-loop control of hazardous chemicals = liability concern, keeps the slider low.
- vertical-ai-orchestration — sits inside the broader mine operations orchestrator.
Open questions
- How does the system learn from genuinely novel ore types without first failing on them at production scale?
- Does the sustainability framing (less energy/reagent) survive lifecycle analysis once the compute cost is included?
Related
- deloitte-ai-dossier-eri — primary source.
- hydrocarbon-reservoir-exploration — parallel pattern, parallel risks.
- synthetic-data-generation — methodology page.
- scaling-wall — the in-source canonical OOD example.