Lance E. Besaw

Counterpropagation Neural Network for Stochastic Conditional Simulation: An Application with Berea Sandstone

A neural network trained using the counterpropagation algorithm to produce stochastic conditional simulations is applied and evaluated on a real dataset. This type of network is a non-parametric clustering algorithm not constrained by assumptions (i.e. normal distributions) and is well suited for risk and uncertainty analysis given spatially auto- correlated data. Detailed geophysical measurements from a slab of Berea sandstone are used to allow comparison with a traditional geostatistical method of producing conditional simulations known as sequential Gaussian simulation. Equiprobable simulations and estimated fields of air permeability are generated using an anisotropic spatial structure extracted from a subset of observation data. Results from the counterpropagation network are statistically similar to the geostatistical methods and original reference fields. The combination of simplicity and computational speed make the method ideally suited for environmental subsurface characterization and other earth science applications with spatially auto- correlated variables.

Parameter Estimation Using an Artificial Neural Network to Incorporate Multiple Types of Data

We apply a modified counterpropagation artificial neural network (ANN) that uses multivariate data to several parameter estimation problems: (1) estimation of small scale Berea sandstone geophysical properties, (2) estimation of apparent conductivity at a leaking landfill using electromagnetic data and (3) estimation of hydraulic conductivity field at a landfill in New York State using pumping test and well log data. The counterpropagation algorithm has been enhanced in this research to allow for spatial interpolation that is comparable to traditional kriging methods. This enhanced ANN is data-driven, can incorporate large amounts of multiple data types to produce parameter estimates in real-time and does not require the computation of large covariance matrices associated with traditional geostatistical methods (kriging).