Combining neural networks with kriging for stochastic reservoir modeling

Stochastic reservoir modeling is being increasingly used for modeling reser voir heterogeneity. This paper describes a methodology to model the distrib ution of reservoir properties using well data and soft geological knowledge in the form of sedimentary and diagenetic patterns. The technique, develop ed based on a combined use of radial basis function (RBF) neural networks a nd geostatistical kriging, is demonstrated with an application to interpola ting porosity in the A'nan Oilfield, located onshore north China. The integ rated technique first uses neural networks to estimate the porosity trends from high-dimensional geological patterns. Optimization of the network perf ormance is done by variogram analysis of the residuals at the conditioning points. Gaussian simulation of the residuals is then performed, and the res ulting residual maps are combined with the porosity trends obtained from ne ural networks. From the case study, the results are realistic and honor the geological rules of the oilfield. The technique is fast and straightforwar d, and provides a computational framework for conditional simulation.