Archived project

Bayesian Sequential Simulation (BSS)

Goal: BSS simulates a primary variable based on a collocated secondary variable. see https://github.com/Raphael-Nussbaumer-PhD/BSS

Updates
0 new
0
Recommendations
0 new
0
Followers
0 new
19
Reads
1 new
393

Project log

Raphaël Nussbaumer
added a research item
Bayesian sequential simulation (BSS) is a geostastistical technique, which uses a secondary variable to guide the stochastic simulation of a primary variable. As such, BSS has proven significant promise for the integration of disparate hydrogeophysical data sets characterized by vastly differing spatial coverage and resolution of the primary and secondary variables. An inherent limitation of BSS is its tendency to underestimate the variance of the simulated fields due to the smooth nature of the secondary variable. Indeed, in its classical form, the method is unable to account for this smoothness because it assumes independence of the secondary variable with regard to neighbouring values of the primary variable. To overcome this limitation, we have modified the Bayesian updating with a log-linear pooling approach, which allows us to account for the inherent interdependence between the primary and the secondary variables by adding exponential weights to the corresponding probabilities. The proposed method is tested on a pertinent synthetic hydrogeophysical data set consisting of surface-based electrical resistivity tomography (ERT) data and local borehole measurements of the hydraulic conductivity. Our results show that, compared to classical BSS, the proposed log-linear pooling method using equal constant weights for the primary and secondary variables enhances the reproduction of the spatial statistics of the stochastic realizations, while maintaining a faithful correspondence with the geophysical data. Significant additional improvements can be achieved by optimizing the choice of these constant weights. We also explore a dynamic adaptation of the weights during the course of the simulation process, which provides valuable insights into the optimal parametrization of the proposed log-linear pooling approach. The results corroborate the strategy of selectively emphasizing the probabilities of the secondary and primary variables at the very beginning and for the remainder of the simulation process, respectively.
Raphaël Nussbaumer
added a project goal
BSS simulates a primary variable based on a collocated secondary variable. see https://github.com/Raphael-Nussbaumer-PhD/BSS
 
Raphaël Nussbaumer
added 4 research items
The geostatistical integration of geophysical and hydrological data has significant potential for adequately characterizing heterogeneous aquifers. Although progress has been made in this regard, the proposed methods were so far mostly limited to the local scale. Recently, Ruggeri et al. (2013) proposed a two-step Bayesian sequential simulation (BSS) approach, which allowed for extending the integration of geophysical and hydrological data to the sub-regional scale. While this approach demonstrated great promise, it is computationally very expensive, which so far prevented its extension to regional-scale datasets. To overcome this problem, we propose to modify the classic BSS algorithm as follows: (1) simulate the grid in a multi-scale manner, (2) use the same visiting path for all simulations to compute the kriging weights only once, and (3) choose kriging neighbourhood based on superblock and quadrant-spiral search.
Bayesian sequential simulation (BSS) is a powerful geostatistical technique, which notably has shown significant potential for the assimilation of datasets that are diverse with regard to the spatial resolution and their relationship. However, these types of applications of BSS require a large number of realizations to adequately explore the solution space and to assess the corresponding uncertainties. Moreover, such simulations generally need to be performed on very fine grids in order to adequately exploit the technique's potential for characterizing heterogeneous environments. Correspondingly, the computational cost of BSS algorithms in their classical form is very high, which so far has limited an effective application of this method to large models and/or vast datasets. In this context, it is also important to note that the inherent assumption regarding the independence of the considered datasets is generally regarded as being too strong in the context of sequential simulation. To alleviate these problems, we have revisited the classical implementation of BSS and incorporated two key features to increase the computational efficiency. The first feature is a combined quadrant spiral – superblock search, which targets run-time savings on large grids and adds flexibility with regard to the selection of neighboring points using equal directional sampling and treating hard data and previously simulated points separately. The second feature is a constant path of simulation, which enhances the efficiency for multiple realizations. We have also modified the aggregation operator to be more flexible with regard to the assumption of independence of the considered datasets. This is achieved through log-linear pooling, which essentially allows for attributing weights to the various data components. Finally, a multi-grid simulating path was created to enforce large-scale variance and to allow for adapting parameters, such as, for example, the log-linear weights or the type of simulation path at various scales. The newly implemented search method for kriging reduces the computational cost from an exponential dependence with regard to the grid size in the original algorithm to a linear relationship, as each neighboring search becomes independent from the grid size. For the considered examples, our results show a sevenfold reduction in run time for each additional realization when a constant simulation path is used. The traditional criticism that constant path techniques introduce a bias to the simulations was explored and our findings do indeed reveal a minor reduction in the diversity of the simulations. This bias can, however, be largely eliminated by changing the path type at different scales through the use of the multi-grid approach. Finally, we show that adapting the aggregation weight at each scale considered in our multi-grid approach allows for reproducing both the variogram and histogram, and the spatial trend of the underlying data.
Bayesian sequential simulation (BSS) is a powerful geostatistical technique, which notably has shown significant potential for the assimilation of datasets that are diverse with regard to the spatial resolution and their relationship. However, these types of applications of BSS require a large number of realizations to adequately explore the solution space and to assess the corresponding uncertainties. Moreover, such simulations generally need to be performed on very fine grids in order to adequately exploit the technique’s potential for characterizing heterogeneous environments. Correspondingly, the computational cost of BSS algorithms in their classical form is very high, which so far has limited an effective application of this method to large models and/or vast datasets. In this context, it is also important to note that the inherent assumption regarding the independence of the considered datasets is generally regarded as being too strong in the context of sequential simulation. To alleviate these problems, we have revisited the classical implementation of BSS and incorporated two key features to increase the computational efficiency. The first feature is a combined quadrant spiral - superblock search, which targets run-time savings on large grids and adds flexibility with regard to the selection of neighboring points using equal directional sampling and treating hard data and previously simulated points separately. The second feature is a constant path of simulation, which enhances the efficiency for multiple realizations. We have also modified the aggregation operator to be more flexible with regard to the assumption of independence of the considered datasets. This is achieved through log-linear pooling, which essentially allows for attributing weights to the various data components. Finally, a multi-grid simulating path was created to enforce large-scale variance and to allow for adapting parameters, such as, for example, the log-linear weights or the type of simulation path at various scales. The newly implemented search method for kriging reduces the computational cost from an exponential dependence with regard to the grid size in the original algorithm to a linear relationship, as each neighboring search becomes independent from the grid size. For the considered examples, our results show a sevenfold reduction in run time for each additional realization when a constant simulation path is used. The traditional criticism that constant path techniques introduce a bias to the simulations was explored and our findings do indeed reveal a minor reduction in the diversity of the simulations. This bias can, however, be largely eliminated by changing the path type at different scales through the use of the multi-grid approach. Finally, we show that adapting the aggregation weight at each scale considered in our multi-grid approach allows for reproducing both the variogram and histogram, and the spatial trend of the underlying data.