Standard Test Models

A Test Everyone Agrees On

An engine is only as trustworthy as the test it passes, and the test has to be shared. If two people run different codes on their own private models, they cannot compare results. So the community settled on a handful of benchmark models, and the most famous is Marmousi: a densely faulted, folded section built specifically to defeat simple imaging. The rule of thumb is blunt: if your migration handles Marmousi, it handles most real data.

We cannot ship the exact Marmousi grid, but its character is easy to reproduce procedurally. This mini-Marmousi has the same ingredients, many dipping layers, a central anticline, and faults that offset the whole stack, all driven by a single complexity knob. Turn the knob up and the flat layer cake acquires the dip, the fold, and the faults that make imaging hard.

Why It Fails Convolution, and Why Resolution Is Free

A convolutional section of this model would be wrong, and instructively so. The faults shed diffractions that convolution cannot make; the steep flanks of the fold need a true wave-equation migration to focus; and the fault blocks cast shadows that distort the layers beneath them. Those are exactly the physics Part 5 showed convolution omits, which is why a model like this is the proving ground that separates a real imaging engine from a fast approximation.

Because the whole model is a formula, resolution is a free choice. The identical structure rasterises onto a coarse grid where an engine runs in seconds for prototyping, or onto a fine grid for the real, expensive test. That is the deep advantage of procedural models over digitised ones: you never have to choose the resolution in advance. The final builder section turns any of these models into a portable file you can hand to a research code, complete with the bytes on the wire.