The geoML manual

Theory, and how it ties to the code. Each chapter teaches one piece of the model in geostatistical terms, ends its theory sections with an In the code box naming the classes that implement them, and carries small runnable examples on the bundled data. The full case studies close the manual.

Derivations live in the papers. They are cited in the text and listed in full in a References section at the end of each chapter that has any. The design documents under docs/ carry the engineering decisions and measurements, and are linked where they apply.

Code follows the package’s own conventions, which the geo-ml skill in plugins/geoml/ records: variables and attributes are reached by tree path (container.values("V/prediction")), never by attribute chain.

Written by Claude. Every code block here runs, and the numbers quoted in the text were measured rather than recalled, but that guarantees only that the code executes and that the figures match what produced them. There may be errors in the geostatistical reasoning, in the emphasis, in the choice of what deserved a chapter, and in claims about why something works. Where the manual and the package disagree, the package is correct; where the manual and the papers disagree, the papers are.

Status: complete draft. Every code block is executed by run_blocks.py, which runs the fenced blocks of a chapter in one namespace, in order:

python docs/manual/run_blocks.py docs/manual/*.md

A full pass takes about 20 minutes on a desktop GPU, most of it in four chapters that train real models: 5 (a deep network), 13 and 15 (a cross-validation each) and 17 (two 3D implicit models). Everything else runs in seconds.

Everything runs on the bundled sample data except chapter 17, which fetches about 20 kB of drillhole logs on first run and caches them in docs/manual/data/; it is skipped, not failed, when there is no connection and nothing cached.

Part I — From kriging to the variational GP

  1. Why another geostatistics

  2. The GP is kriging

  3. Inducing points and the ELBO

  4. Warpings, likelihoods, and where the Gaussian sits

  5. Latent networks

  6. Categories and boundaries

  7. Simulation

  8. Change of support

Part II — The workflow

  1. From database to data

  2. Containers and addressing

  3. Building and training

  4. Prediction, blocks and surfaces

  5. Validation

  6. Reporting

Part III — Case studies

  1. Walker Lake

  2. Jura

  3. A folded quartz vein in 3D