QI workflow: from seismic to rock properties

Why QI is a chain

QI is a sequence of stages, each computed from the output of the one before:

• Data: amplitude-preserving angle stacks (in Figure 7.1 a near stack at 8° and a far stack at 32°) and well logs of $V_{\mathrm{P}}$, $V_{\mathrm{S}}$ and density.
• Well tie and wavelet: a synthetic made from the logs is matched to the seismic at the well, and the match fixes the wavelet: its timing, its spectrum and its phase.
• Inversion: the stacks are inverted with that wavelet for P- and S-impedance, $Z_{\mathrm{P}}$ and $Z_{\mathrm{S}}$, and so for $V_{\mathrm{P}}/V_{\mathrm{S}}$. The data hold no frequencies below a few hertz, so the inversion needs a low-frequency model, usually built from wells and mapped horizons, for the absolute level of each property.
• Rock physics: a rock-physics model calibrated to the logs predicts where each rock and fluid plots in ($Z_{\mathrm{P}}$, $V_{\mathrm{P}}/V_{\mathrm{S}}$), the rock-physics template of Section 7.2. It is the dry frame that is calibrated; Gassmann’s equation (Section 5.3) then adds the fluid.
• Classification: every inverted sample is given a probability of being each facies, here shale, brine sand or gas sand, and a porosity, by comparing it with the template.
• Reservoir: the classes and porosities become maps of net pay and hydrocarbon pore volume, the inputs to volumetrics, simulation and the choice of where to drill.

Each stage consumes the previous stage’s output, so it cannot tell an error made upstream from the earth: it builds on the error as if it were geology. Each stage also needs different expertise (processing geophysicist, petrophysicist, rock physicist, inversion specialist, reservoir modeller), which is why QI is a team’s work and why the hand-offs between stages are where errors go unnoticed.

Exercise, follow an error down the chain

The figure opens with 20% noise in the stacks, the true wavelet and the well’s low frequencies carried along the horizon. The headline sentence and the table measure every stage against the earth, beside the best case, which has no noise, the true wavelet and the earth’s own low frequencies.

• Read the plates in order, (a) to (h): the earth, the far stack computed from it, the wavelet and the tie at the well, the inverted impedance, the template, the classified section and the gas pore thickness. Each is computed from the one before it.
• Set the low-frequency model to None. The data cannot supply the impedance’s absolute level, so inside the thick sand the inverted impedance drifts back toward the mean in (e), and the template reads much of that sand as shale in (g): the sand cells called shale rise from 125 to 269, the gas pore thickness goes from 12% to 26% off, and its total from 0.1% to 26% low.
• Set Well again and the noise to 0. The tie is perfect, yet 364 shale cells are called sand, 165 of them beyond the pinch-out and most of the rest where the sand thins: the low-frequency model carries the well’s sand along the horizon into ground where there is none, and the classifier, whose spreads were measured at the well where that model is right, is too sure of itself everywhere else. The invented sand lies below the contact and is called brine, so the gas map escapes (5% off, as in the best case) while a net-sand map would not.
• Return the noise to 20% and rotate the wavelet by 45°. The tie’s correlation falls only to 0.71, a value often accepted on real data, but the impedance error grows from 4.6 to 6.0%, 76% of the reservoir cells take the wrong class and the gas pore thickness is 22% off, its total 18% low.
• Return the phase to 0° and raise the noise to 40%. The tie still reads 0.97 and the line’s total comes out only 4% low, yet the gas pore thickness in (h) is 21% off: gas invented in one place and gas missed in another cancel in the total, not in the map. The headline is the map error for that reason.
• Finally choose Earth, no noise and the true wavelet: the best case. What remains, 26% of the reservoir cells in the wrong class but only 5% error in the gas pore thickness, is not the bandwidth alone. Most of the wrong cells are porous brine sand at the left end called shale: brine sand there sits close to shale in the template, and the inversion assumes a linear reflectivity that the exact data do not obey. With data made by that linear law the share falls to about 16%, and even a perfect inversion leaves 7%, at cells the sand’s top or base crosses.

How QI integrates Parts 3 to 6

QI is not a new body of knowledge. It joins the parts of this book you have already worked through.

• Part 3 (structural framework) supplies the geometry the QI volumes sit in. The low-frequency model is built along mapped horizons (Section 3.4) and respects the faults (Section 3.2); traps identified in Section 3.6 become QI targets.
• Part 4 (stratigraphy) supplies the depositional context. The sequence-stratigraphic surfaces of Section 4.2 are the horizon framework, the depositional systems of Section 4.3 constrain which rocks to expect where, and seismic geomorphology (Section 4.6) checks the spatial pattern of QI results against plausible depositional shapes.
• Part 5 (rock physics and AVO) supplies the physics. Gassmann fluid substitution (Section 5.3) is the fluid step of the rock-physics model, the AVO theory of Sections 5.4 and 5.5 is what a pre-stack inversion inverts, and the synthetic seismogram of Section 5.6 is the well tie.
• Part 6 (attributes) supplies further inputs and checks. Spectral decomposition helps isolate tuning, coherence masks faults, and attribute-based classification (Section 6.5) is a cousin of the probabilistic classification QI uses.

The step from interpreter as picker to interpreter as modeller is mostly a matter of extending what Parts 3 and 4 taught into a quantitative framework, not of abandoning it.

Where QI projects fail

QI looks clean on paper. In practice, projects fail for well-understood reasons, most of which Figure 7.1 can reproduce:

• Bad input data. Gathers whose amplitudes were not preserved, a poor velocity model, too few wells, or wells that all sample one facies. No later stage can recover what the data do not hold.
• A wrong rock-physics model. If the template does not match the logs (anisotropy, microcracks, cement or unusual mineralogy that the model ignores), every prediction downstream of it is biased.
• A bad low-frequency model. The band-limited data leave the absolute impedance to the model. Without one the result is relative impedance; with one taken from too few wells, the wells’ rocks are carried into places they are not (the second step of the exercise).
• Unpropagated uncertainty. Deterministic maps presented as truth when the data support only probabilities. Note that a classifier calibrated at the well where everything matches will understate its own uncertainty away from it.
• Over-claimed resolution. Vertical resolution is limited by the seismic bandwidth to about a quarter of a wavelength, roughly 8 to 40 m at reservoir depths. Detail finer than that is not known, however sharp the map looks.
• Weak integration with the asset team. A QI product that sits in a folder and does not inform where wells are drilled has failed, however good its numbers.

Deliverables and deliverable formats

A typical QI project ends with the following set of deliverables, each produced by specific stages of the pipeline:

• 3D elastic volumes (from stage 3): $Z_{\mathrm{P}}$, $Z_{\mathrm{S}}$, $V_{\mathrm{P}}/V_{\mathrm{S}}$ and density. Standard SEG-Y or proprietary volume format, same geometry as input seismic.
• 3D rock-property volumes (from stage 4): porosity, shale volume, saturation. Same format.
• 3D facies probability volumes (from stage 5): one probability volume per facies class. Same format.
• Reservoir property maps (horizon-based): porosity-thickness maps, net-to-gross maps, hydrocarbon-saturated pore volume (HCPV) maps. These drive volumetrics.
• Uncertainty quantification: P10/P50/P90 maps of key reservoir properties. Risk maps integrating multiple uncertainty sources.
• Calibration report: well tie quality, RPM calibration match, synthetic-to-real seismic match, cross-validation with blind wells.
• Decision deliverable: prioritized well-location recommendations, drainage maps for field development, production-forecast inputs for reservoir simulation.

Every deliverable is traceable back to the stage that produced it, so the asset team can understand what uncertainty affects what prediction.