Prediction of Fluid Velocity in Highly Heterogeneous Conductivity Fields Using a Genetic Algorithm-Designed Artificial Neural Network

A genetic algorithm (GA) is used to select the operational parameters of artificial neural networks (ANN) which are trained to predict fluid velocity. Populations of three-layer, feedforward backpropagation ANN's with varying numbers of hidden nodes, types and slopes of activation functions, alpha and beta learning rates and initial distributions of weights for both the input and hidden layers are created by the GA. The GA- defined ANN's are trained with inputs-output pairs of hydraulic conductivity neighborhoods and resulting fluid velocities at certain points in the simulation domain. The hydraulic conductivity fields are highly heterogeneous with an ensemble log conductivity variance of 1.0. Results of the GA are defined and selected ANN velocity predictions are presented.