Sequential Gaussian Simulation

Key Takeaways

• The variogram is the fundamental geostatistical tool that controls the spatial structure reproduced by SGS, describing how the spatial correlation between property values changes as a function of the separation distance between measurement locations: the experimental variogram is computed from the available hard data (well log measurements) by calculating the average squared difference between pairs of data values at increasing separation lags, producing a variogram cloud that shows how rapidly property values become uncorrelated with increasing distance; a theoretical variogram model (spherical, exponential, or Gaussian mathematical form) is fitted to the experimental variogram to provide a smooth, continuous description of the spatial correlation for use in the kriging estimation step of SGS; the key variogram parameters include the range (the distance at which the variogram reaches its sill value, representing the maximum correlation length beyond which property values are spatially uncorrelated), the sill (the maximum variogram value, equal to the variance of the property for a stationary data set), and the nugget (the apparent variogram value at zero lag, representing the combined effect of measurement error and sub-scale variability below the sampling resolution); vertical and horizontal variograms are typically different (the correlation range is usually much longer horizontally, along depositional layers, than vertically, across layer boundaries), and the variogram anisotropy ratios (horizontal-to-vertical range ratios of 10:1 to 1000:1 are common) must be correctly characterized to reproduce the layered nature of reservoir heterogeneity that controls fluid flow; a poorly fitted variogram produces SGS realizations that either over-smooth the property distribution (if the range is too long) or generate artificial short-range variability (if the range is too short), both of which distort the fluid flow behavior predicted by the resulting reservoir simulation model.
• Normal score transformation is the preprocessing step that converts the original petrophysical data (which may have a skewed, lognormal, or bimodal distribution) to a standard Gaussian (normal) distribution with mean zero and variance one before SGS is applied, because the algorithm requires that the simulated property follow a multivariate Gaussian distribution: the normal score transform ranks each data value from smallest to largest, assigns a standard normal variate to each value based on its rank in the data set, and records the transform function; after simulation in the Gaussian space (where the SGS algorithm is exact), the simulated values are back-transformed to the original data space using the inverse transform, reproducing the original histogram of the property; if the data has multiple modes (a bimodal distribution of high and low permeability corresponding to two distinct rock types), the normal score transform converts this to a single mode in the Gaussian space, potentially losing the bimodal structure that is geologically meaningful; in practice, bimodal petrophysical data requires a different approach (truncated Gaussian simulation or plurigaussian simulation) that explicitly models the discrete rock type distribution before simulating the continuous petrophysical property within each rock type, rather than transforming the combined multimodal distribution to a single Gaussian.
• The relationship between porosity and permeability SGS realizations is a critical modeling challenge because porosity and permeability are correlated (high-porosity zones tend to have high permeability), but separate SGS runs of each property without co-conditioning would produce realizations where the porosity and permeability spatial distributions are independent — not correlated as they are in the real reservoir: co-simulation algorithms (co-sequential Gaussian simulation or co-kriging-based SGS) are used to simultaneously simulate two or more correlated properties, using the cross-variogram between properties in addition to the direct variograms to ensure that the simulated spatial patterns respect the inter-property correlation; alternatively, porosity can be simulated first (as the primary variable with more data, usually from porosity logs in wells and seismic impedance transform), and permeability can then be co-simulated conditioned on both the permeability data (from core plug measurements) and the simulated porosity field (using the empirical porosity-permeability regression relationship as the co-conditioning constraint); the co-simulation approach ensures that the permeability realizations honor both the direct variogram of permeability and the spatial distribution of porosity, producing a more geologically realistic representation of the permeability field than independent simulation of each property.
• Uncertainty quantification using multiple SGS realizations forms the foundation of probabilistic reserves estimation and development decision analysis in reservoir characterization: a typical uncertainty study uses 20-100 SGS realizations (each with a different random seed but all conditioned to the same well and seismic data), runs each realization through a reservoir flow simulator (often simplified from the full-physics model for computational efficiency), and compares the resulting production forecasts from all realizations to build a probability distribution of the expected production performance; the P10, P50, and P90 of the resulting production forecast distribution (corresponding to the 10th, 50th, and 90th percentiles of the simulated outcomes) are used to define the proved (P90, conservative), probable (P50, median), and possible (P10, optimistic) reserves categories reported to regulatory agencies and stock exchanges; the range of the production forecast distribution (P10 minus P90 divided by P50) provides a measure of the uncertainty in the reserve estimate that is attributable to inter-well spatial heterogeneity, complementing the structural uncertainty (trap volume uncertainty) and fluid contact uncertainty that are handled separately; in development decisions where the choice between different well counts, spacing patterns, or completion strategies depends on the actual reservoir heterogeneity, the distribution of production outcomes from SGS-based uncertainty analysis provides the probabilistic basis for decision tree analysis of the development options.
• Seismic-constrained SGS uses secondary information from 3D seismic data (typically acoustic impedance derived from seismic inversion, or seismic amplitude attributes correlated with porosity or fluid content) to constrain the SGS realizations away from wells and to improve the spatial resolution of the reservoir model between well locations: in seismic-constrained SGS, the seismic attribute is first converted to a soft probability or a seismic-derived porosity estimate using the well-calibrated correlation between the seismic attribute and the petrophysical property, then used as a secondary variable in the co-simulation or as a trend model (a spatially variable mean) that biases the conditional probability distribution in the direction of the seismic-derived estimate; the result is a set of SGS realizations that reproduce the fine-scale (well-log resolution) variability of porosity and permeability from the variogram model while also honoring the lateral trends in the seismic data that reflect the large-scale heterogeneity of the reservoir (channel bodies, carbonate build-ups, diagenetic fronts) that the variogram model alone cannot capture; the seismic constraint dramatically reduces the uncertainty in the SGS realizations compared to well-only conditioning, because the dense spatial sampling of seismic data provides information between wells that constrains the inter-well property distribution more tightly than the well data alone; however, the accuracy of seismic constraint depends on the quality of the seismic data and the strength of the well-calibrated seismic-petrophysical relationship, and a poorly calibrated seismic constraint can introduce systematic bias into all realizations that worsens the reservoir model accuracy.