Live AI Lab / Mineral processing, metallurgy & R&D teams
Mining AI Laboratory — reagent recipe optimisation
Learn from decades of flotation test records, then let a surrogate model choose the next lab test for a new ore composite — and see how many tests it saves.
Before you start
- Aurora Metallurgical Laboratory (fictional) is fictional. Its 529 historical flotation tests (reagent-archive-2026.11.1) and every “lab test” in this demo come from a documented simulator. No company's data is used, and the demo is not connected to any plant.
- The model, its validation and the search are computed for real on our server each time you press Run. The step-by-step animation is a replay of that recorded computation, paced so you can follow it.
- You can run up to 6 studies per session. Please don't enter personal information.
This check helps keep the demo available and affordable for everyone.
Ore-processing digital twin
Before you start
- This is a fictional nickel/copper concentrator (“Polaris Concentrator”) with synthetic data only. It is not Nornickel data and is not connected to any plant or control system.
- The balance is a transparent illustration, not a calibrated plant prediction. Every figure is a stated assumption or is computed from one — see the AI demo notice.
- You can run the balance up to 6 times per session. Please don't enter personal information.
This check helps keep the demo available and affordable for everyone.
Want this on your own process?
We can build a balance against your plant's calibrated recovery and grade model, with a process engineer reviewing every assumption. Tell us your ore system and we'll prepare a tailored, non-binding example.
The request goes to a real, monitored review process — a person reads it. Nothing is purchased or committed.
How it works
The problem
Every new ore composite needs its reagent scheme re-screened in the lab. One-factor-at-a-time screening takes many tests, and decades of earlier results sit unused in reports.
What it does
- A Gaussian-process surrogate trained on the archive and validated on a later-years hold-out against a baseline
- Constrained Bayesian optimisation proposes the next test with its uncertainty and probability of meeting grade and MgO limits
- A step-by-step replay of the computed run, with real compute time shown separately from playback
- A deterministic, conserved nickel/copper mass balance for the flowsheet view
Steps
- Pick an ore composite (or mix your own) and set grade, MgO and dose limits
- The model reviews the archive, selects its kernel, validates on later runs and scores every recipe
- Each recommended recipe is tested in the simulated lab and the model updates, compared with one-factor-at-a-time screening
Guardrails
- Synthetic archive and simulated lab tests — no real company data, and never presented as real experiments
- Recommendations are experiments for a metallurgist to review, not dosing or plant-control settings
- Never connects to a plant or control system; reagents are generic classes, not product formulations
This demonstrates a workflow using synthetic data. Results are examples, not outcomes measured at any organisation. A pilot measures results on your own approved data.