01 / INPUTS
Assemble the evidence
Known locations, geological domains, routine assays, returned BBWi tests and campaign constraints.
GEOMETALLURGICAL DRILLHOLE OPTIMISER / SYNTHETIC DEMONSTRATION
Which tests will
teach us the most?
You have geological domains, routine assays and existing drill core. Metallurgical testwork is limited. How do you choose the next batch of tests to learn more about the targets that matter?
This demonstration explores that decision using Bond Ball Work Index (BBWi), mutual information, geological context and explicit cost constraints. The objective is not simply to sample where variance is highest, but to choose the affordable batch whose possible results are expected to tell us the most about weighted mine-plan targets.
Opens the app on Posit Connect Cloud. All example data are synthetic. This is not mine validation, a guarantee of savings or a recommendation to execute a real testwork programme.
Evidence → model → selection → learning.
02 / MODELS
Represent uncertainty
Compare assay-based and spatial models while keeping prediction checks separate from the decision objective.
03 / INFORMATION
Measure what each test can teach us
Use mutual information in bits to quantify how observing a proposed batch is expected to contract uncertainty at weighted mine-plan targets, while discounting redundant tests.
04 / LEARNING
Review and update
Inspect the shortlist, costs and diagnostics; simulated returned results illustrate how evidence changes.
What to look for in the app.
Start with the defaults and run the demonstration. Follow the existing-core sample-selection example first. The important question is why these intervals were selected, not simply whether a plot looks impressive.
Locations and domains: compare returned samples, eligible intervals and selections. A geographically isolated sample is not automatically the most informative sample for the decision targets.
Information and cost: the acquisition objective is mutual information in bits. For the Gaussian model, one bit corresponds to a two-fold contraction of the relevant target uncertainty volume. Expected variance reduction is shown separately as an intuitive consequence, not substituted for the information objective.
Redundancy matters: two tests can each look useful on their own but contain overlapping information. The optimiser evaluates the batch jointly, so the value of the batch is not simply the sum of individual scores.
Prediction checks: read held-out error, interval coverage and interval width together. A fitted-versus-observed plot alone does not establish out-of-sample accuracy. Models sharing the same samples are not competing sampling strategies.
Information theory, made practical.
A high-variance location is not automatically the best place to spend testwork budget. The useful question is how much a possible observation changes what is known about the targets that matter. That is why the optimiser uses target-weighted mutual information and then applies campaign cost and feasibility constraints.
The interactive app includes a short How it works section covering the Gaussian process response model, virtual conditioning, entropy / mutual information, joint-batch redundancy, cost constraints and the return–refit–repeat workflow.
The value is a decision you can explain.
Choosing the next test is a geological, statistical and operational problem. The useful output is an inspectable recommendation: which intervals, under which assumptions, at what modelled cost, and with what expected information.
This example shows how I connect domain knowledge, uncertainty-aware modelling and practical software. A site-specific application would begin with the actual decision, data quality, sample representativity and constraints—not by assuming the demonstration settings transfer unchanged.
Scope and limits.
This landing page focuses on laboratory test selection from already-drilled core. It does not claim validated optimisation of new drill-hole locations. Synthetic response relationships are assumptions, and model-conditional uncertainty is not proof of improved real-world accuracy.
The interactive app is public; its source repository remains private. Please do not send confidential datasets in an initial enquiry.