GEOMETALLURGICAL TESTWORK OPTIMISER / FICTIONAL RESEARCH BETA
Which available core should we test next?
A good testwork programme has to answer more than “which sample looks interesting?” We need geological coverage, suitable physical specimens, a credible budget and a reason to believe another test might improve what we know.
The interactive DH Optimiser illustrates a two-stage approach: first establish feasible geological coverage, then use Bayesian predictive Information Gain to choose further tests as laboratory results become available.
Runs as a public Shiny application on Posit Connect Cloud. Begin with the fictional presets. This is testwork selection from already-drilled core, not designing new drilling collar locations.
TWO CONNECTED STAGES
From material availability to model learning.
STAGE A / SPECIMEN PLANNING
Can we test it, and what geology is missing?
Assess declared core inventory, laboratory requirements, availability, remaining sample mass, shared stock, geological domains and coverage. The planner proposes an affordable response-blind coverage programme.
Try the fictional D1 → B1 example: two proposed tests cost AUD 80 and AUD 140, exhausting the AUD 220 budget. No material is actually purchased and neither test is assumed to have returned a BBWi result.
A real inventory needs owner and laboratory verification. Unverified stock remains unknown or conditional, not automatically eligible.
STAGE B / SIMULATED LABORATORY
What did the results teach us?
Start with returned fictional BBWi observations, propose a coverage specimen, make a simulated purchase, reveal only that selected measurement and refit the observed-only Gaussian-process (GP) model.
After the coverage requirement is met, the existing information-gain method can propose another eligible and affordable test. The interface shows charged cost, expected uncertainty reduction and post-result model uncertainty separately.
This is a separate fictional scenario from D1/B1; it has its own budget and frozen response oracle.
Which physical specimens are eligible and what geological domains have valid, returned testwork?
02 / PLAN
Fill coverage gaps
Select affordable specimens using only information available before new tests return.
03 / RETURN
Reveal selected BBWi
Purchase a fictional test, return its simulated response and update the observed-only model.
04 / LEARN
Choose another test
Score remaining feasible candidates by the expected reduction in model uncertainty, then examine what actually changed.
The Shiny app also includes Try your data: a bounded educational upload route with a downloadable fictional CSV and predictive-model comparisons. The public service is not approved for confidential geological, laboratory or client records.
WHY INFORMATION GAIN?
Expected learning depends on the target.
The GP estimates covariance between possible observations and geological prediction targets. The simplest question is: how much should conditioning on a candidate reduce uncertainty about those targets?
IGw(S) = ½ Σt wt log₂(vt / v′t|S)
This is a weighted sum of marginal target entropy reductions under the fitted model—not joint target-vector mutual information, economic Expected Value of Sample Information (EVSI), or a guarantee that future observed prediction error will be smaller.
Two individually informative tests may overlap in what they teach the model. The selected combination also needs to respect laboratory costs and physical constraints. The fictional examples show those mechanics, without establishing policy superiority or operational calibration.
Important limitations. The BIF geology, laboratory outcomes, sample-stock declarations and prices used in the demonstrations are fictional. The GP suitability state is LIMITED; its uncertainty is not certified as calibrated for real mining decisions. Proposal costs and expected Information Gain are not realised economic savings.
Before using real inventory, verify specimen custody, available material, preparation, laboratory method and cost conventions. Do not upload confidential, proprietary, licensed or personal information to the public demonstration.
Ready to explore the complete testwork workflow?
Start with the fictional examples, trace each selection and inspect the model warnings and downloadable audit. For a real programme, the geological population, laboratory requirements and decision objective need to be defined and checked first.
