What survives

A rock has many independent properties. A seismic amplitude has one number, or two where offset is available. The difference has to go somewhere, and the earlier modules located it. The card on the left collects those losses, with the size of each one measured rather than described.

The losses are of different kinds. Some are collapses: several numbers becoming one, as when velocity and density become an impedance. Some are saturations: a property that stops responding once it has changed a little, as gas saturation does. Some are blurs: the wavelet in time, the noise in amplitude. And one is a confusion: half the reflection belongs to a rock that was never the subject of the question.

Only the blurs get better with better data. The collapses and the saturation are properties of the physics and no amount of acquisition or processing touches them.

Module 07 counted the rocks consistent with a measured amplitude, using a tolerance picked to represent the noise. Now that tolerance can be computed properly: module 08 gives the uncertainty in a fitted intercept from the angle range, the fold and the noise, and the set of admissible rocks is everything within two of those standard deviations.

The panel shows the survivors as a population — what porosities, what fluids, what saturations. The count is large, and the important thing about it is not the number itself but which of its dimensions are wide and which are narrow.

Changing the noise widens or narrows the whole set, and changing the mute does it again. Neither of them changes the geology by so much as a percent of porosity.

Now require the gradient to match as well, within its own two standard deviations. The surviving set shrinks. How much it shrinks is the practical case for doing AVO, stated as a figure rather than as a conviction.

At moderate noise the gradient removes 55% of what the intercept left. That is a real and useful return. At high noise it removes 25%, because the gradient's error bar has grown so wide that it barely constrains anything, which is module 08's result in a form that can be costed.

Then look at which dimensions shrank. Vp/Vs narrows sharply, which is exactly what module 05 predicted: offset measures shear. Porosity narrows a little. And saturation does not move at all. The second measurement constrains lithology and fluid type, and not how much is there.

Every trap below comes from a rule that works most of the time, which is why each has survived in use: a rule that failed often would have been discarded. Each button loads the situation and the analysis that goes with it.

The statement below is assembled from the analysis on the previous steps. Each clause is backed by a count or a range from the surviving models: nothing in it is asserted that the numbers do not support, and nothing that they do support is left out.

Changing the noise, the mute or the fold changes the wording with them. What a statement can claim depends on how well the measurement was made, so it is re-derived for each set of conditions rather than carried over.

The shape of it is consistent. There is usually something firm to say about impedance and about the presence of a hydrocarbon, usually nothing to say about saturation, and almost always a caveat about the seal, because half of every amplitude belongs to a rock nobody characterized.

The point of the whole set is not that AVO is unreliable. It is that AVO answers some questions well, some badly and some not at all, and that which is which can be worked out in advance rather than discovered after a well.

• Ask the question the data can answer. "Is there a hydrocarbon here" is often answerable. "How much" usually is not. Framing a prospect around the second one guarantees a disappointing result no matter how good the processing is.
• Compute the value of the gradient before paying for it. Step 3 puts a number on what the second measurement removes. At high noise it removes a fifth of the solutions, which may not justify the work.
• Spend on what narrows the set. Step 5 ranks the candidates. Very often the cheapest useful measurement is a constraint on the seal rather than anything about the reservoir.
• Report a set, not a value. Every number in a quantitative interpretation is a representative of a set of possibilities. Quoting the representative without the set is the single most common way these methods mislead.

The practice that follows from all of this is to look for the other models once one has been found that explains the data. The first fit reads as an answer, and it is one member of a set whose size has not yet been checked. Everything in these ten modules is an elaboration of that.

This site is itself a model, and inherits the limitation it describes. The rock physics is one choice among several, the noise is simpler than real noise, nothing here is spatial, and a real interpreter has structural conformance, well control and a hundred other things this page does not. The numbers describe the parameter space that was searched, which is not the earth.

Everything here is built from the earlier modules. The rock physics is module 01 and 02's, the reflection coefficients are exact Zoeppritz as in module 05, the intercept and gradient are fitted as in module 06, and their uncertainties come from module 08's closed form. Each Method tab lists its own assumptions and all of them apply here.

The tolerances are computed, not chosen. A model is kept if its intercept is within two standard deviations of the measured one, and — in step 3 — its gradient likewise. Those standard deviations follow from the angle range, the number of traces and the noise level on the sliders, so moving any of those genuinely changes what is admissible rather than just relabelling it.

The search grid is porosity 5–35% in 1% steps, clay 0–30% in 10% steps, fluid brine, oil or gas, hydrocarbon saturation 10–100% in 10% steps, and shale velocity 2300–3100 m/s in 200 m/s steps, with a hydrocarbon at zero saturation counted once as brine rather than three times. The fit for each model uses eight angles rather than sixteen, which changes the recovered gradient by about 0.001 and makes the sweep fast enough to run while a slider moves.

Counts are not probabilities. This is the same caution as module 07's and it matters more here, because the closing statement is generated from the counts. The survivor count is the number of cells in a hand-chosen grid, weighted equally, that reproduce the measurement. Turning it into a probability needs a prior that reflects the basin, the depositional setting and the trap, and different priors give materially different answers from identical seismic. Read every count as "this many distinct explanations exist", which is a statement about the measurement, and never as "this is how likely each is", which is a statement about geology that seismic alone cannot make.

Where this departs from production practice.

• Nothing is spatial. This is the largest omission on the page. Real interpretation gains enormously from an anomaly conforming to structure and repeating over an area, and none of that is available to a single-gather analysis.
• No well control, no seismic inversion, no other attributes, all of which narrow the set further and are the reason quantitative interpretation is worth doing at all.
• The truth is known to the page and hidden from the analysis. That is a luxury: on real data nobody can check the answer, which makes the discipline of counting alternatives more important rather than less.
• Thickness is fixed at 45 m and tuning is therefore not part of the search here, though module 07 showed it widening the admissible set considerably.
• The generated statement is a teaching device. It is assembled from the counts by a fixed set of rules. It is not a report, and a real one would carry the geological reasoning that this page has no access to.