AAPG Annual Convention and Exhibition

Datapages, Inc.Print this page

Iterative Stochastic Calibration of Basin Scale Migration Models


Monte Carlo simulation techniques have been used before in the modelling of oil and gas migration at basin scales. Here we discuss how results from such simulations are used in an iterative approach to improve a basin scale geologic model. The input parameters to the basin simulator are iteratively conditioned by selecting those simulation runs that show low misfits between modelled and observed oil and gas column heights in the observation wells. The selected low misfit simulation runs will frequently constitute less than 1% of the total number of simulation runs in order to obtain good results from the analysis. The results from many simulation runs cannot be used because they do not provide a good enough match to the observation data. Input model parameters are described as stochastic (normal) distributions. The input properties for each layer must be treated as independent of each other. Model parameters that are routinely adjusted after the iteration include source rock parameters (TOC, HI, thicknesses), lithology fractions (shale, carbonate, sand fractions) and hydraulic flow properties (permeabilities, entry pressures, capillary pressures). In addition, depth horizons of many basins are uncertain in the source rock kitchen areas, and laterally variable depth dependent standard deviations may therefore be applied to describe the uncertainties of the depth model. New misfit-weighted estimates of the mean and standard deviation of all input parameters for all layers are estimated after the iteration. These are used as input distributions to the next iteration. The iterative loop can be stopped when the estimates of the mean and standard deviations remain similar between iterations, e.g. the when all the means change less than 0.3 standard deviation. After the final iteration run has been completed, the estimate of the standard deviation may still be improved by simply running many more simulation runs, thus increasing the number of low-misfit runs. We demonstrate the approach with data from a North Sea migration study where more than 30.000 simulation runs were used to compile the a-posteriori probability distributions for the input parameters. In this case study, the TOC and HI properties for some of the layers – that had first been derived from geochemical data in the basin – had to be adjusted in order to obtain an optimum stochastic calibration of trapped oil and gas in the calibration wells.