Guillaume Pirot

A practical guide to performing multiple-point statistical simulations with the Direct Sampling algorithm

The Direct Sampling (DS) algorithm is a recently developed multiple-point statistical simulation technique. It directly scans the training image (TI) for a given data event instead of storing the training probability values in a catalogue prior to simulation. By using distances between the given data events and the TI patterns, DS allows to simulate categorical, continuous and multivariate problems. Benefiting from the wide spectrum of potential applications of DS, requires understanding of the user-defined input parameters. Therefore, we list the most important parameters and assess their impact on the generated simulations. Real case TIs are used, including an image of ice-wedge polygons, a marble slice and snow crystals, all three as continuous and categorical images. We also use a 3D categorical TI representing a block of concrete to demonstrate the capacity of DS to generate 3D simulations. First, a quantitative sensitivity analysis is conducted on the three parameters balancing simulation quality and CPU time: the acceptance threshold t, the fraction of TI to scan f and the number of neighbors n. Next to a visual inspection of the generated simulations, the performance is analyzed in terms of speed of calculation and quality of pattern reproduction. Whereas decreasing the CPU time by influencing t and n is at the expense of simulation quality, reducing the scanned fraction of the TI allows substantial computational gains without degrading the quality as long as the TI contains enough reproducible patterns. We also illustrate the quality improvement resulting from post-processing and the potential of DS to simulate bivariate problems and to honor conditioning data. We report a comprehensive guide to performing multiple-point statistical simulations with the DS algorithm and provide recommendations on how to set the input parameters appropriately.

A pseudo genetic model of coarse braided‐river deposits

A new method is proposed to produce three-dimensional facies models of braided-river aquifers based on analog data. The algorithm consists of two steps. The first step involves building the main geological units. The production of the principal inner structures of the aquifer is achieved by stacking Multiple-Point-Statistics simulations of successive topographies, thus mimicking the major successive flooding events responsible for the erosion and deposition of sediments. The second step of the algorithm consists of generating fine scale heterogeneity within the main geological units. These smaller-scale structures are generated by mimicking the trough-filling process occurring in braided rivers; the imitation of the physical processes relies on the local topography and on a local approximation of the flow. This produces realistic cross-stratified sediments, comparable to what can be observed in outcrops. The three main input parameters of the algorithm offer control over the proportions, the continuity and the dimensions of the deposits. Calibration of these parameters does not require invasive field measurements and can rely partly on analog data.

Probabilistic inversion with graph cuts: Application to the Boise Hydrogeophysical Research Site

Inversion methods that build on multiple-point statistics tools offer the possibility to obtain model realizations that are not only in agreement with field data, but also with conceptual geological models that are represented by training images. A recent inversion approach based on patch-based geostatistical resimulation using graph cuts outperforms state-of-the-art multiple-point statistics methods when applied to synthetic inversion examples featuring continuous and discontinuous property fields. Applications of multiple-point statistics tools to field data are challenging due to inevitable discrepancies between actual subsurface structure and the assumptions made in deriving the training image. We introduce several amendments to the original graph cut inversion algorithm and present a first-ever field application by addressing porosity estimation at the Boise Hydrogeophysical Research Site, Boise, Idaho. We consider both a classical multi-Gaussian and an outcrop-based prior model (training image) that are in agreement with available porosity data. When conditioning to available crosshole ground-penetrating radar data using Markov chain Monte Carlo, we find that the posterior realizations honor overall both the characteristics of the prior models and the geophysical data. The porosity field is inverted jointly with the measurement error and the petrophysical parameters that link dielectric permittivity to porosity. Even though the multi-Gaussian prior model leads to posterior realizations with higher likelihoods, the outcrop-based prior model shows better convergence. In addition, it offers geologically more realistic posterior realizations and it better preserves the full porosity range of the prior.

Towards 3D Probabilistic Inversion with Graphcuts

For appropriate uncertainty quantification in hydrogeological applications (e.g., contaminant plume forecasting), it is essential to infer subsurface models that feature geologically realistic geometries and property contrasts. Recently, an efficient multiple-point statistics probabilistic inversion approach, with model proposals based on graph cuts, has been shown to provide posterior model realizations that honor pre-defined geometrical shapes and property contrasts. It has been tested for both synthetic and field examples involving crosshole ground-penetrating radar. Here, we present the approach and proceed with initial findings on how to extend this method to 3D and hydraulic tomography data. Improvements and modifications in the Markov chain Monte Carlo algorithm are proposed that allow for appropriate acceptance and convergence rates. We also discuss possible ways to circumvent long computing times, for example, by including physical approximations and machine learning techniques, or to focus on global optimization rather than Bayesian posterior sampling.