Geological modeling (geobrain.geomodel)

Everything that builds an earth before any physics touches it: geostatistics, implicit geological modelling, and the frames that carry the result.

The distinction the module is built around

Kriging gives you the best single map: minimum error variance, and therefore smoother than the earth. Simulation gives you maps with the right variability, each one wrong in a different way. Which you want depends on the question: an average grade wants the estimate; a probability of exceeding a cutoff wants the ensemble, because you cannot read a probability off a smooth map.

Geostatistics (geobrain.geomodel.geostats)

Category	What is there
Kriging	SimpleKriging, OrdinaryKriging, UniversalKriging, BlockKriging, IndicatorKriging, SoftIndicatorKriging, CollocatedCokriging
Simulation	SGSIM, DSSIM, SISIM, CoSGSIM, LUSIM, FFTMA, PlurigaussianSim
Multiple-point	SNESIM, FILTERSIM, DirectSampling, ImageQuilting
Generative	VAESimulator, GANSimulator, DiffusionSimulator
Variograms	VariogramCalculator, CrossVariogramCalculator, CovarianceModel, VariogramKernel, automatic fitting with diagnostics
Transforms	NormalScore, BoxCox, YeoJohnson, Detrend, Decluster, IndicatorEncode
Validation	KFold, LeaveOneOut, PostSimulation, PostIndicatorKriging

A field, start to finish:

import numpy as np
from geobrain.geomodel import (FFTMA, CovarianceModel, GeoGrid,
                               PropertyMetadata, SimulationExecutionConfig,
                               VariogramKernel)

grid = GeoGrid(shape=(60, 40), origin=(0.0, 0.0), spacing=(10.0, 10.0))
model = CovarianceModel(nugget=0.0, structures=[
    VariogramKernel(kind=VariogramKernel.SPHERICAL, contribution=1.0,
                    ranges=(150.0, 60.0, 1.0e6), angles=(0.0, 0.0, 0.0))])

sim = FFTMA(model,
            property=PropertyMetadata(name="V", kind="continuous", unit="1"),
            execution=SimulationExecutionConfig(n_realizations=1, seed=7))
frame = sim(None, grid).realizations[0].frame
values = np.asarray(frame.to_numpy("simulation"), dtype=float).reshape(-1)
print(values.shape, f"mean {values.mean():+.3f}, sd {values.std():.3f}")

(2400,) mean -0.115, sd 1.078

Warning

The variogram ranges bind to the grid axes positionally, and a flat field unflattens with x varying fastest, so values.reshape(NY, NX) is [y, x], ready for imshow. Reshaping the other way round and transposing gives an earth that still looks plausible and has its anisotropy mirrored. If your sample locations stop sitting on the features they were drawn from, this is why.

Frames and models

Object	Carries
GeoGrid	a regular grid domain: shape, origin, spacing
GeoPoints	scattered locations
GeoFrame	geometry plus named properties; what a kriging or simulation call takes and returns
PropertyMetadata	a property’s name, kind (continuous / categorical) and unit
EarthModel / Field	several coupled properties on one mesh, which a joint inversion updates together
Category	a named class with a code, because a facies is a name, not a number

Implicit modelling (geobrain.geomodel.implicit)

ImplicitModel builds geometry from what a geologist actually records, namely surface points, orientations, series and faults, as the level sets of a scalar field. It is differentiable with respect to the input data, so the derivative of a unit’s volume with respect to a contact point’s position is one backward pass, and a block model can be an inversion parameter rather than a fixed input.

SurfacePointData, OrientationData, SeriesDefinition, FaultDefinition and ImplicitModelConfig are the inputs; StackRelation says how series stack.

A study end to end, from a sample table to the probability a decision needs. From examples/02_geomodel/07_case_study.py.

See also

• examples/02_geomodel/01_estimation_vs_simulation.py: why the two answer different questions, measured on the same data.

• examples/02_geomodel/03_kriging_estimators.py: what each estimator assumes about the mean.

• examples/02_geomodel/06_implicit_modelling.py: geometry that carries gradients.

• examples/02_geomodel/07_case_study.py: decluster, transform, fit, simulate, and finish with a probability rather than a map.