Faults in the Model
Learning objectives
- Cut a layered model with a normal or reverse fault
- See reflectors step across the fault by the throw
- Explain why a clean offset makes good training labels
- Notice that convolution makes no fault-edge diffraction
Cutting the Section
Real earth is faulted, and a fault is the simplest structural discontinuity: a plane across which one block has slipped relative to the other. Add one to the layered model and every reflector on the moved side shifts by the same amount, the throw. Push the throw positive for a normal fault, negative for a reverse fault, and slide the fault across the section.
Clean Offsets and a Missing Diffraction
Two things are worth seeing at once. First, the offset is clean. In a 2D convolutional section the reflectors simply jump across the fault by the throw, with sharp, unambiguous geometry. That cleanliness is a feature: it is exactly what makes convolution the tool of choice for generating fault-detection training data, because the label, the fault plane you drew, is known perfectly and the section is cheap to produce by the thousand. The next section makes exactly that dataset.
Second, look along the fault plane itself. There is no fan of energy spreading from the fault tip. A real fault edge is a sharp scatterer that radiates a diffraction, one of the strongest cues an interpreter uses to place a fault, and one of the features fault-detection networks key on in real data. Convolution makes none of it, because it has no lateral wave propagation. So a convolutional fault section is geometrically truthful and physically incomplete in a specific, nameable way. Holding both of those in mind, useful yet limited, is the judgement this whole part is training. Part 5 restores the diffraction with the wave equation.