Insights / Metallurgy and process development
What a surrogate model can and can't learn from flotation tests
Most mineral processing groups hold years of flotation test results in reports and spreadsheets. A surrogate model can be trained on those records to suggest the next test for a new composite. Whether it deserves any lab time depends on how it is validated, and on being clear about what it cannot do.
What it can do
A Gaussian-process surrogate estimates recovery and grade for a recipe and ore, along with its uncertainty. That uncertainty is what makes it useful for test selection. It can favour tests that look promising or would teach the model the most, within limits on grade, impurities and reagent dose, and each new result updates it.
Validate on later tests, not a random split
A random train and test split lets a model learn from tests run in the same campaign, on the same ore, by the same people as the tests it is scored on. That flatters it. Holding out the most recent years is closer to the real question: could this have helped on work that had not happened yet? Compare the result with a simple baseline, such as nearest similar tests. If the model doesn't beat the baseline, stop there.
Where historical data lets you down
- Assay methods, laboratories and procedures change over time, and the records may not say so
- Failed or abandoned tests are often missing, which biases what the model sees as possible
- Below-detection values and missing assays need explicit handling, not a quiet zero
- Reagents renamed or reformulated over the years have to be reconciled by someone who knows them
- Different ore types need their own check. A model trained on sulphide tests says little about a spodumene circuit
What it can't do
A surrogate fitted to bench tests is not a plant model. It doesn't set dosing, and it doesn't predict concentrator recovery without separate work to relate lab and plant results. Claims about tests saved or recovery gained mean nothing until they are measured on your own data, under a protocol your metallurgists agreed before seeing the answer.
How to find out whether your archive is usable
- Export the tests with dates, composite or ore identifiers, reagents, doses and assays
- Record which runs were excluded and why, rather than deleting them
- Agree the time-based hold-out and the baseline before any model is fitted
- Have a metallurgist review each proposed test, and decide in advance what result would end the work