Insights / Metallurgy and process development
How a surrogate model picks the next flotation test
This is what our Mining AI Laboratory does when you run a study, including the numbers it uses. The archive and lab results are synthetic, from a fictional laboratory, so the point is the method and not the outcome.
1. Read the archive and record what was left out
The study starts by reviewing the historical tests. Excluded records are shown with the reason, so nothing disappears quietly, and the page lists the historical ores nearest to the chosen composite.
2. Choose the model settings from the data
The surrogate is a Gaussian process. How quickly it believes results change with each input is a setting that must be chosen, and so is how much noise it assumes. The demo scores 162 combinations of these settings by log marginal likelihood and keeps the best. This costs some compute, so the fitted model is cached for each archive version, and the page shows real compute time separately from the replay.
3. Check it on later years
The model is fitted on tests up to a cut-off year and scored on tests from 2020 onward, which it never saw. The same split is scored for a baseline: the average of the five most similar historical tests. Showing both matters, because a model that can't beat that baseline has not earned a say in lab planning.
4. Score every recipe within your limits
You set the grade, MgO and reagent dose limits. Each candidate recipe within the dose caps gets a predicted recovery, an uncertainty, and a probability of meeting the limits. The score is expected improvement multiplied by that probability, so a recipe needs both a realistic chance of beating the best so far and a realistic chance of meeting specification.
5. Test, update and compare
The top recipe goes to a simulated lab test, the result updates the model, and the loop repeats. The same simulator runs a one-factor-at-a-time screen, so you can compare the two sequences side by side.
What this does and doesn't show
It shows the mechanics, and that each step can be inspected. It doesn't show that the approach will work on your ore, because a simulator is far tidier than a laboratory. Recommendations are experiments for a metallurgist to review, not dosing or plant settings. On real work, the same steps would run on your archive, with acceptance criteria agreed first.