From logged drillholes to grade shells, block models, and grade–tonnage reports — with the parameters, inputs, and identity behind every number carried on the result itself.
The problem
Mineral resource work is judged on defensibility. A block model that cannot say which samples informed it, which variogram was used, who ran it, and whether it can be reproduced is a picture, not evidence. GeoMine Scientist runs the estimation and modelling chain on a separately certified numerical foundation and attaches that record to every artifact it produces.
30+
Governed compute kernels — the whole chain to a resource statement
3
Epistemic classes every artifact declares: observation, estimate, interpretation
4
Refusal categories, each with its own action
0
Client-side numerics — nothing is recomputed to draw
The resource modelling chain
Each stage consumes the artifact the last one produced, and every result names the inputs it came from.
Log & desurvey
Collars, surveys, logged intervals, and assays come from the governed database, and traces are computed from surveyed attitude — refusing to continue past the deepest survey station.
Screen (QA/QC)
Assay eligibility resolves each batch and interval from reference materials, blanks, duplicates, and checks — with every threshold a declared input and none defaulted.
Composite & decluster
Assays are composited to a declared support inside a single domain, and declustering weights correct clustered sampling instead of biasing the mean.
Model & domain
Implicit fields fit logged contacts, isosurfaces turn them into watertight shells, and estimation domains are read against the interval log.
Characterise
Top-cut evidence, variography, and a declared block-model lattice are assembled — evidence a Competent Person selects over, never a silent default.
Estimate & simulate
Ordinary kriging returns an estimate and variance per block inside a declared search, and sequential Gaussian simulation produces a realization stack.
Validate
Kriging neighbourhood analysis, cross-validation, and swath plots show the estimate against the data — and density gives every block a tonne.
Classify & report
A classification policy assigns a category from the run's own evidence, and the resource statement reports tonnage, grade, and metal by category and cut-off.
What the mining module does
Geometry, estimation, uncertainty, reporting, assumptions, and the record behind all of it.
Desurvey logged holes, model contacts implicitly, and contour the result into nested grade or boundary shells — then read all of it in one 3D scene with sections, vertical exaggeration, a ruler, and synced-camera comparison.
Krige a block model with an anisotropic variogram and a declared search neighbourhood, then run conditional simulation to see the spread the single estimate hides.
Grade–tonnage curves carry a P10–P90 band across realizations, contained metal is computed per realization rather than multiplied from two means, and the report says plainly that it is a quantity and not a resource classification.
Each result is identified by the hash of its own content. Its manifest records the kernel and its digest, the input identities, the parameter hash, the randomness root, the acting identity, and the toolchain pins.
Parameters that decide the answer — the variogram, the tail models, the search neighbourhood, the stationarity mean — are required fields with no pre-filled value, and each explains why no default exists.
When a job is invalid the system refuses with a code, a category, and the stage it failed at — four categories, each demanding a different response — rather than returning a plausible number nobody can defend.
Inside the workflow
Screens from the working application — the same evidence a technical reviewer would be handed.
Collars, downhole surveys, logged intervals, and assays come from the governed project database rather than a spreadsheet export. Desurveying computes each trace from surveyed attitude — and refuses to continue past the deepest survey station, because carrying on down the last known bearing is an assumption about ground nobody measured. Traces render in 3D coloured by the unit logged at each station, with nothing interpolated between them.
Logged contacts become anchor points, an implicit scalar field fits a declared stratigraphic column through them, and isosurfaces contour that field into watertight shells. Nested shells contoured from the same field render as layers of one model, ordered by isovalue, each carrying its own recorded volume, surface area, and Euler characteristic — and an open stratigraphic surface is declared open rather than reported with a meaningless volume.
Ordinary kriging returns an estimate and its kriging variance per block, inside a search neighbourhood the job had to declare — a block with no sample in range is refused, not quietly interpolated from data the variogram says is uncorrelated. Sequential Gaussian simulation then produces a stack of equally probable realizations, and an exceedance model turns that stack into the probability each block clears a stated cut-off.
Grade–tonnage curves report tonnage above cut-off as P10, P50, and P90 across realizations. Contained metal is computed per realization and then averaged, because multiplying the reported mean tonnage by the reported mean grade gives a different — and wrong — number. And the report says plainly that it states a quantity, not a resource classification: classification under a reporting code is a Competent Person's judgement, and no category label is emitted here.
The foundation
The science does not live in the browser. GeoMine Scientist consumes a separate, governed compute platform through a machine-readable contract — the kernel registry, the job structure, the artifact manifests, and the refusal vocabulary — and performs no client-side numerics of its own. That platform now spans the full chain: 30+ governed kernels carry a deposit from desurveying and QA/QC through estimation and simulation to classification and a resource statement, each stage a content-addressed artifact.
Kernels declare a determinism class and an execution class, and a job whose declared policy disagrees with the kernel's is refused before anything runs.
The same job returns the same artifact identity — numerical behaviour is a contract down to the order of floating-point operations, and worker count or restart changes nothing.
Stored artifacts are content-addressed and re-verified when retrieved — a tampered byte or a dangling reference is an error, not a silent answer.
A Rust engine and an independent Python reference are held bit-exact against each other, so the numbers are a checked agreement rather than one implementation's word.
The mining estate
The resource model is the centre of gravity, not the whole of it. Exploration geochemistry, pit planning, grade control, mine dewatering, and tailings are their own applications — each governed, each on the same drillholes, projects, and records.
See it end to end
A maiden resource, a profitable pit, a wet cutback, a mill-versus-model gap, a tailings review — each carried across the modules as one thread of evidence, with the value stated plainly.
Beyond the resource model
A resource model is one stage. Ground engineering, mine water, lease-wide monitoring, and field operations run on the same governed foundation — so the drillholes behind a block model and the boreholes behind a pit slope are one dataset, not two.
Go deeper
Each area has its own page with the capabilities, screens, and outcomes behind it.
Mining runs on the same governed foundation as the GIS, groundwater, and geotechnical modules — so the drillholes behind a resource model and the ground investigation behind a pit wall are the same estate, not two systems.
Mining FAQ
Straight answers on scope, classification, reproducibility, and where the geologist's judgement stays.
GeoMine Scientist is the mining module of the SpatialTechSolutions geoscience platform. It covers the resource-modelling chain — drillhole desurveying, compositing, implicit geological modelling, ordinary kriging, sequential Gaussian simulation, exceedance probability, and grade–tonnage reporting — and attaches full provenance to every result.
Not on its own. The chain now runs all the way to a classification and a resource statement — tonnage, grade, and contained metal by category, domain, and cut-off. But the classifier computes a category from the run's own evidence and approves nothing, the resource kernel asserts no conformance to any reporting code, and every statement it produces is provisional until a Competent Person signs it. It makes the reading reproducible; it does not replace the judgement or the sign-off.
Every artifact is identified by the hash of its own content, and its manifest records the kernel and digest, input identities, parameter hash, randomness root, acting identity, and toolchain pins. Re-running the same job returns the same identity, so a reviewer can confirm a figure rather than take it on trust.
Parameters marked as assumptions — the variogram model, the tail models, the search neighbourhood, the stationarity mean — decide the reported answer. A pre-filled dropdown would be a decision nobody actually made, so those fields are required, empty, and each explains why no default exists.
Each modelled element needs its own variogram and its own reading of continuity. Additional assayed elements are carried through the data chain, and each is modelled deliberately rather than by stamping one continuity model across all of them.
No. The geological interpretation — the stratigraphic column, contact polarity, anisotropy orientation, and classification — is supplied by the geologist and recorded as their claim, including the identity of the person who ran the job. The platform makes that reading geometric and reproducible; it does not invent it.
Request demo
Walk through desurveying, implicit modelling, kriging, simulation, and grade–tonnage reporting on a deposit your team already knows — and see what the provenance record looks like at the end of it.