Gaussian Techniques

Key Takeaways

• The normal score transform (also called the Gaussian anamorphosis or the probability integral transform) is the foundational preprocessing step in all Gaussian simulation techniques that converts the original data histogram (which is typically log-normal or skewed for permeability, or bounded between 0 and 1 for porosity and saturation) to a standard Gaussian distribution with mean zero and variance one, by replacing each data value with the z-score of its corresponding quantile in the standard normal distribution: the quantile of each data value is computed from the empirical cumulative distribution function of the data, and the corresponding z-score is found from the inverse normal CDF (the probit function); after simulation in the Gaussian space using the standard Gaussian variogram model, the simulated values are back-transformed to the original property units using the inverse normal score transform, reproducing the original data histogram in each realization; the back-transformation is performed by linear interpolation in a table of (z-score, original value) pairs constructed from the empirical CDF, with extrapolation to extreme values outside the data range performed by a tail model (Gaussian, power law, or constant) that must be chosen to represent the geological expectation for values beyond the sampled range.
• Sequential Gaussian simulation (SGS) generates realizations by visiting each unsampled grid cell in a random sequence and drawing a simulated value from the local conditional Gaussian distribution at each cell: the conditional distribution at each unsampled location is characterized by the conditional mean (the simple kriging estimate of the property value at that location using the conditioning data -- original well data plus all previously simulated values) and the conditional variance (the simple kriging variance, representing the residual uncertainty after conditioning); a random value drawn from the Gaussian distribution with this mean and variance is assigned to the current cell and added to the conditioning data for subsequent simulation steps; this sequential procedure ensures that the simulated field reproduces the input variogram model (the spatial correlation structure of the data) in addition to the Gaussian histogram, because the kriging system used at each step uses the variogram to determine the weights assigned to the conditioning data; the random number seed controls which of the infinitely many possible Gaussian random fields consistent with the data and variogram is realized, and different seeds produce different realizations that together quantify the geological uncertainty remaining after conditioning to the available well data.
• Gaussian co-simulation extends sequential Gaussian simulation to incorporate seismic or other secondary data by performing the simulation in the framework of collocated cokriging, conditioning each simulated value not just on the primary data and previously simulated values in the neighborhood but also on the secondary (seismic) value at the simulation location: in the normal score domain, the collocated cokriging system uses the primary variogram, the secondary variogram (computed from the seismic attribute), and the cross-variogram (computed from the co-located primary-secondary pairs at well locations) to weight the contribution of the seismic value to the conditional mean at each grid cell; the resulting simulations are constrained by the seismic in a way that reproduces the observed correlation between the primary property and the seismic attribute at well locations, using the seismic as a spatially dense guide to interpolation between the sparse wells while preserving the stochastic variability consistent with the residual uncertainty given the seismic constraint; Gaussian co-simulation is used routinely in deepwater reservoir characterization to build porosity and facies models constrained by seismic acoustic impedance, providing better spatial control of reservoir property distribution than well-only simulation while quantifying the residual uncertainty arising from the imperfect correlation between seismic impedance and reservoir porosity.
• Limitations of Gaussian techniques in facies-controlled reservoirs arise from the assumption of univariate Gaussianity and the use of two-point variogram statistics: real geological heterogeneity in fluvial, deltaic, and turbidite reservoirs is not adequately described by a Gaussian variogram model because the spatial patterns of sand and shale (or channel and overbank, or amalgamated and isolated turbidite) are geometrically complex (curvilinear channels, lenticular sand bodies with asymmetric cross-sections, interconnected lobes) and strongly non-Gaussian (the property distribution is bimodal, with sand-proportion peaks at both high and low porosity corresponding to the two facies), requiring either indicator simulation methods (which model the spatial distribution of facies indicators as correlated 0-1 variables) or object-based simulation methods (which simulate the geometry of individual geological objects -- channels, lobes, crevasse splays -- and fill them with Gaussian property distributions) to reproduce the spatial connectivity and geometric character of the geological facies; the practical question of whether Gaussian or non-Gaussian methods are needed for a specific reservoir depends on how strongly the heterogeneity is facies-controlled (if most porosity and permeability variability occurs within facies rather than between facies, Gaussian techniques may be adequate even in geologically complex systems).
• Uncertainty quantification using Gaussian simulation ensembles is performed by generating 50 to 200 realizations of the reservoir property distribution (using different random seeds but the same variogram model and conditioning data) and propagating each realization through the reservoir simulation to produce an ensemble of production forecasts that spans the range of possible reservoir configurations consistent with the available geological data: the P10, P50, and P90 production forecasts (the 10th, 50th, and 90th percentile of the ensemble production at each forecast time) quantify the low, mid, and high cases used in economic evaluation, development planning, and reserve classification; the spread of the ensemble (the distance between P10 and P90) represents the geological uncertainty in the reservoir model that cannot be resolved with the current data, and the reduction of that spread through additional appraisal drilling, seismic acquisition, or pressure surveillance provides the quantitative basis for ranking the value of information from different data acquisition strategies; this uncertainty quantification framework, enabled by the computational tractability of Gaussian simulation methods, has become the standard approach to probabilistic reserve estimation and development optimization in the petroleum industry.