Sondara, Drillhole Estimation and Simulation Compared on Real Held-Out Holes
Published:
Drillhole estimates are usually judged by the method that produced them. The question I wanted answered is a different one: which method predicts a hole that never entered the model, on real data, and is the difference larger than the noise between holes? Sondara is built to answer it, and to say so when a model learned nothing.
Real sources, the same targets
Rocklea Dome (CSIRO, 5,035 one-metre multielement intervals in 158 holes), Alberta MAR_19860002 (22 inclined holes with logged geology) and NTGS 12LE002 (one hole with eleven measured survey stations) are pinned by hash and run through eight stages: acquire, ingest, preprocess, dataset, features, train, infer, evaluate. Splits are by whole hole, inside the drilled area and at its margins, and every method predicts the same targets.
Twelve methods, one scoring
Nearest neighbour, inverse distance, simple, ordinary and universal kriging, LMC cokriging, indicator kriging and sequential Gaussian simulation run on geocond, the engine published on PyPI from its own repository (0.8.0); SNESIM (MPSlib) and Direct Sampling simulate the Alberta logs under two labelled training images; DeepKriging and KCN train on the GPU and ship as audited ONNX models. Paired hole-block intervals decide whether a difference is real, and shuffled-label controls and a training-mean reference say whether a model learned anything.
What it showed on Rocklea iron, 23 test holes: ordinary kriging at RMSE 13.53 wt%, simple kriging, inverse distance and cokriging within 0.4 wt% of its error, DeepKriging at 14.11 and KCN at 14.56, both better than their own controls by 2.3 to 5.4 wt% of error and neither better than kriging. Sequential Gaussian simulation keeps the variance that kriging halves. On Alberta’s hole-group split, with three test holes, no method beats the training mean.
The data it came with
The collection also ships a hyperspectral export. Read against its own files, its hem/goe column is the wavelength of an iron-oxide absorption in nanometres, not a ratio; its assay-like columns copy the assay workbook; and twelve holes have spectral depths offset from the assays. Calibrated on training holes, the iron-oxide index predicts iron with RMSE 11.99 wt% on 701 test rows.
Not a resource estimate
Predictions are values at the sample centres, not block grades, and the simulations are conditional on interpretations that are labelled as such. The live page still shows the earlier local-first viewer; the web product rebuilt on these results comes after the export and validation stages. Version 0.12.000, in construction.
