ACS PRF | ACS
All e-Annual Reports
41520-AC8
Simulation of a Hydrocarbon Reservoir Analog using a Process-Based Model of Fluvial Stratigraphy
The research objective was to use a new and simplified forward fluvial stratigraphy model to simulate the distribution of channel and overbank deposits in the Rhine-Meuse delta, as seen in borehole data. The model will be conditioned such that the simulated stratigraphy fits the borehole data within specified tolerance limits, and the fitting of the model to the borehole data will use optimization techniques of random search and genetic algorithms.
The new 3D fluvial stratigraphy model was developed incorporating key processes only, with a very short run time of 0.5 seconds. This short run time was required since the optimization involves up to 1 million model runs per scenario. Genetic algorithm software was developed so as to be tightly coupled to the stratigraphy model to allow for minimal run times, easy model output evaluation and scenario preparation. A wide range of visualization procedures were programmed including plots to analyze the evolution of the genetic algorithm and 3D visualizations of observed and modeled stratigraphy.
A test area of 15 x 15 km2 was selected in the centre of the Rhine-Meuse delta, the Netherlands, containing Holocene fluvial deposits with approximately 15 individual channel belts and a depositional thickness of approximately 8 m. The fluvial stratigraphy model was fitted to sets of 5, 10, 15, 20 and 25 boreholes, randomly selected from approximately 400 borehole descriptions available in the test area. Fitting the model involved the optimization of a goal function that quantifies the difference in observed and modeled vertical position of channel belt sandstone bodies. Fitting using the genetic algorithm was done by adjusting 12 model inputs, some of which were adjusted for each individual time step. Beforehand, optimal settings of the genetic algorithm were selected using an artificial data set.
Full conditioning of the model to borehole data was possible using up to 20 boreholes. With a larger number of boreholes, the model still fitted borehole data very well, but 100% conditioning was not reached with the computer resources available. By executing multiple optimization runs per data set, several possible model outcomes were found with a complete fit to the data. Using these outcomes, the predicted 3D stratigraphy was compared with that reconstructed from the 400 borehole descriptions. This showed that the fraction of the area with a correspondence between modeled and reconstructed stratigraphy increases with the number of borehole data used for conditioning. With 5 boreholes the model prediction is correct in 40% of the area, with 20 boreholes it is correct in 63% of the area.
The project results are a substantial development in the use of process based models for predicting 3D alluvial stratigraphy. The project shows that the procedure works with: (1) real world data, and; (2) using a number of borehole data commonly available in studies of aquifers or hydrocarbon reservoirs. Also, the project has a large spin-off for fluvial sedimentology modeling. The software for running the genetic algorithm is written in a generic way such that it can be easily used for optimization of other spatio-temporal models. It will be made available to the public as open source through the PCRaster internet site (http://pcraster.geo.uu.nl). Also, the project resulted in improved algorithms for storing modeled erosion and deposition in a 3D block of data.
