14. Reporting

The last mile: turning a validated model into things people look at. The package draws every figure twice, deliberately. A matplotlib family (Explorer) is for print, and a plotly family (Interactive) carries the same figures under the same names for looking at, plus a dashboard that links them. The arithmetic behind both lives in one backend-free module, so the printed figure and the interactive one can never quietly disagree.

The running example is Jura’s seven elements, scored where reporting should score them: on the held-out validation set.

import os
import tempfile

import geoml
import numpy as np

os.makedirs("figures", exist_ok=True)
geoml.set_seed(1234)

jura_train, jura_validation = geoml.datasets.jura()
n_el = len(jura_train.get("Elements").labels)

root = geoml.latent.BasicInput(
    inducing_points=jura_train,
    transform=geoml.transform.Isotropic(1.0))

gp = geoml.latent.BasicGP(
    root,
    size=n_el,
    kernel=geoml.kernels.Spherical())

# one likelihood serves any number of columns; the warping's size is what
# says how many there are
warping = geoml.warping.ChainedWarping(
    geoml.warping.Softplus(n_el),
    geoml.warping.ZScore(n_el))

model = geoml.models.VGPNetwork(
    jura_train, "Elements",
    geoml.likelihood.Gaussian(warping),
    gp,
    options=geoml.models.GPOptions(verbose=False))

model.train_full(max_iter=300)

model.predict(jura_validation, n_sim=30, include_noise=True)

14.1 The honest figures

Chapter 13 made the point, and here it pays off. jura_validation was never trained on, so every comparison figure on it means what it shows. That includes accuracy, which asks the model for measurement intervals, and may be used here without the caveat it carries on training data.

explore = geoml.plots.Explorer(jura_validation, continuous="Elements",
                               model=model)

figure = explore.accuracy()
figure.savefig("figures/14-accuracy.png", dpi=150, bbox_inches="tight")

figure = explore.prediction_scatter()
figure.savefig("figures/14-scatter.png", dpi=150, bbox_inches="tight")

The accuracy plot is the geostatistician’s calibration check, and chapter 13 computes its number as goodness. The scatter is the figure everyone asks for first and over-reads: its spread mixes model error with assay noise, which is exactly why the accuracy plot sits beside it.

14.2 The dashboard

The plotly twins compose into a single self-contained HTML page with their selections linked, so clicking a population in the histogram lights up the same samples on the map. It renders in a notebook through an iframe and travels as one file that needs nothing installed.

board = geoml.plots.Interactive(
    jura_validation,
    continuous="Elements",
    categorical="Rock").dashboard(figures=("histogram", "scene"),
                                  plotlyjs="cdn")

path = board.write_html(
    os.path.join(tempfile.mkdtemp(), "jura-dashboard.html"))

print("dashboard written:", os.path.getsize(path) > 10000)

plotlyjs="cdn" keeps the file small by loading the plotting library from the web. The default embeds it instead, for a page that works with no connection at all, which is the right choice for a report that must open in ten years.

14.3 What goes in a report, and where it came from

A defensible reporting set, chapter by chapter, all from one model:

item	source
block model with grades and their doubt	chapters 8, 12 (as_data_frame, as_pyvista, to_zarr)
grade–tonnage with uncertainty	chapter 8’s figure
domain boundaries and grade shells as meshes	chapter 12 (get_contour, DXF)
validation scores and calibration	chapter 13 (cross_validate, conformalize)
calibration and scatter figures	this chapter, on held-out data
the model document itself	print(model) and to_dot() (chapters 5, 11)

The last row deserves more custom than it gets. The model’s own repr, with every parameter, every bound and what was fixed, is the audit trail a competent person signs against, and it costs one line to include.

In the code. plots/explorer.py and plots/interactive.py hold the twin families, plots/prepare.py the shared arithmetic, and plots/dashboard.py the linked page. as_pyvista(...) on any container is the door to external 3D viewers.

Further reading

Chapter 15 assembles the full sequence, data to dashboard, on Walker Lake, and chapter 17 does it with the geometry of a real deposit in the way.