Fewer lab tests to find the right reagent scheme.
Every new ore composite means re-screening reagents in the lab — often one factor at a time, over months. Decades of earlier test results usually sit unused in reports. We turn that archive into a model that proposes the next most informative test, with its uncertainty, for your metallurgists to approve.
How it works
- Read the archive
Historical flotation tests are cleaned, versioned and checked: duplicates, excluded runs and missing assays are recorded, not hidden.
- Train and validate
A surrogate model learns how ore and recipe drive recovery, grade and impurities — and is tested on later campaigns it never saw, against a simple baseline.
- Propose the next test
Every recipe within your dose and quality limits is scored for expected improvement and the chance of meeting specification.
- Learn from each result
Each lab result updates the model immediately. Your metallurgists approve every proposed test.
What the demo shows
The Mining AI Laboratory runs the full pipeline for real on every click — kernel selection, hold-out validation, scoring thousands of recipes and a test-by-test optimisation loop — then replays the recorded computation step by step.
It also runs a conventional one-factor-at-a-time screen on the same simulated lab, so you can compare how many tests each approach needs.
Honest scope: the demo uses a synthetic archive and a documented simulator from a fictional laboratory. It contains no company's data, and the numbers it shows are not a measured saving. A pilot measures results on your own archive.
What a pilot needs
- An export of historical lab tests (any format; we do the cleaning)
- The current reagent scheme and its limits
- A metallurgist who reviews proposed tests
- Agreement on where the data is hosted and who can access it
What you get
- A validated model with its error on your own later campaigns
- Ranked next tests with uncertainty and feasibility
- A written report you keep, whatever you decide next
What we don't do
The model proposes experiments for people to review. It does not set dosing or connect to plant control systems, and we don't present simulated results as plant outcomes.