Random Field Representation

Is any real aquifer actually homogeneous?

What you’re watching A map of the hydraulic conductivity measured at the Borden research aquifer in Canada — one of the most densely sampled aquifers in the world. Even this clean, relatively uniform sandy aquifer shows conductivity varying from point to point across the field.

The mechanism Heterogeneity is the rule, not the exception. Borden is famous precisely because it is so unusually uniform — one of the most homogeneous natural aquifers ever characterized, with an ln K variance of only about 0.3 (well inside the mildly-heterogeneous band) and a horizontal correlation scale of a few meters. Yet even this exceptionally clean sand varies enough to control how a tracer plume spreads.

Why it matters If the cleanest aquifer on record is this variable, every practical site is worse — which is the honest starting point for groundwater modeling.

IGW-NET Loading a real measured field like Borden into the digital laboratory lets you run flow and transport through actual heterogeneity, not an idealization — the messy truth, simulated and seen.

How do we turn a cloud of conductivity measurements into something we can model?

What you’re watching A second view of the Borden conductivity data, emphasizing the statistical structure — the spread of values and how nearby points resemble one another.

The mechanism Because K spans orders of magnitude, we work with its logarithm, ln K. Taking the log tames a multiplicative, wildly ranging quantity into a well-behaved, roughly bell-shaped (Gaussian) one whose statistics — mean, variance, and spatial correlation — are stable and meaningful.

Why it matters Every stochastic method downstream — random fields, geostatistics, Monte Carlo — rests on this single move of working with ln K.

IGW-NET Seeing the raw K data and its ln K transform side by side shows why the log is the natural language of aquifer heterogeneity.

If we only have scattered measurements, how do we picture the whole aquifer?

What you’re watching Several generated conductivity fields (realizations), each consistent with the same statistics but differing in the exact placement of high- and low-K zones.

The mechanism A random field model fills the gaps between measurements by generating plausible conductivity fields that all honor the same statistics — mean, variance, and correlation. Each realization is one equally-likely version of the aquifer we can never fully measure.

Why it matters These realizations are the raw material for everything stochastic: run flow and transport through many of them and you get a probabilistic, honest prediction.

IGW-NET Hand IGW-NET a mean, a variance, and the correlation scales and it generates a complete random conductivity field instantly — realization after realization on the fly, making ‘many possible aquifers’ concrete and powering the Monte Carlo and probabilistic-capture work.

What does the ‘correlation scale’ of an aquifer actually look like?

What you’re watching Four conductivity fields with different correlation structures: a fine isotropic speckle; a horizontally elongated pattern; strong horizontal layering; and vertical streaking — the same statistics, different spatial scales and orientations.

The mechanism The correlation scale (λ) is the typical distance over which conductivity stays similar — the size of the heterogeneity. Small λ gives a fine speckle; large λ gives big coherent patches. When λₓ = λᵧ the field is isotropic, with no preferred direction; when they differ it is anisotropic — layered along the long-correlation direction.

Why it matters Correlation scale and its anisotropy control channeling, connectivity, and how transport behaves — yet they are far harder to measure than the mean, and routinely under-characterized.

IGW-NET Dialing λₓ and λᵧ and watching the field morph from speckle to layers to streaks turns an abstract geostatistical parameter into a picture you can read at a glance.

How variable is variable — and how do we quantify it?

What you’re watching Conductivity fields at increasing ln K variance: low variance shows gentle, muted contrasts; high variance shows violent swings between very high and very low conductivity.

The mechanism The ln K variance measures the magnitude of heterogeneity — how far conductivity swings from its mean. Low variance is a nearly uniform aquifer; high variance means high- and low-K zones differ by orders of magnitude, producing strong channeling.

Why it matters Variance is the single biggest control on how much heterogeneity distorts flow and transport — and on whether simple effective models can be trusted. Knowing roughly which band a site falls in tells you how much trouble to expect.

IGW-NET Sweeping the variance and watching the field go from bland to violent makes the parameter tangible — and sets up exactly when effective models break down.

Is one statistical model enough to describe every aquifer?

What you’re watching Conductivity fields generated by different random-field models — a smooth multi-Gaussian field versus fields with connected channels or discrete facies — all sharing the same variance and correlation but structurally different.

The mechanism The multi-Gaussian model is the default, but it tends to disconnect the extremes. Real aquifers often have connected high-K channels — paleochannels, fractures — that demand non-Gaussian or indicator/facies models. Same two-point statistics, very different connectivity.

Why it matters Connectivity controls early arrival and channeling; a Gaussian model can badly underpredict the fast transport a channel model captures — a model choice that changes the answer.

IGW-NET Comparing fields from different models, then running transport through each, shows that ‘same statistics’ does not mean ‘same behavior.’

How does a computer actually conjure a realistic aquifer?

What you’re watching Conductivity fields produced by different generation algorithms — each a method for drawing a random field with prescribed statistics.

The mechanism Several algorithms generate correlated random fields: spectral (FFT-based) methods, turning bands, sequential Gaussian simulation, and others. Each imposes the target mean, variance, and correlation by a different numerical route, with trade-offs in speed, accuracy, and the structures it reproduces.

Why it matters Practical stochastic modeling depends on generating many fields fast; the algorithm choice is what makes large Monte Carlo studies feasible.

IGW-NET Generating fields by different algorithms in real time demystifies what is usually a black box — you watch the random aquifer being built.

How do measurements pin down which aquifer is real?

What you’re watching Four panels: the ‘truth,’ an unconditional simulation (right statistics but wrong placement of features), and conditional simulations honoring 100 and 20 measurement points — the 100-point field closely reproduces the truth, the 20-point field only near its data.

The mechanism The crucial point: the real aquifer is not truly random — it is fixed and, in principle, knowable. The randomness represents our lack of data, not genuine indeterminacy in the world. Unconditional simulations honor only the statistics, so their high- and low-K zones land in the wrong places. Conditional simulations are forced to match the measured values at sampled locations, so they honor both the statistics and the data — reproducing the truth near measurements and diverging only between them.

Why it matters This is the honest way to characterize a site: uncertainty that shrinks exactly where you have looked. It also makes ‘data worth’ visible — how much each new measurement reduces uncertainty — which is the basis for deciding where to drill next.

IGW-NET Adding conditioning points and watching the realizations snap toward the truth turns the abstract value of data into something you see happen — the basis for optimal monitoring design.