Neural network approximation of numeric subsurface models in combinational optimization

The ANN-GA approach to design optimization integrates two well-known computational technologies, artificial neural networks (AN%) and the genetic algorithm (GA), with a simple scheme for exploiting a network of common workstations to reduce the computational burden associated with applying formal optimization techniques to subsurface engineering problems. The greatest computational investment in a design project of the kind which will be described in this paper is in the simulation of physical processes needed to calculate the cost function. The ANN-GA methodology addresses this problem by training ANNs to stand in for the simulator during the course of a search directed by the GA. The ANNs are trained and tested from examples stored in a reusable knowledge base of representative simulations which relate variations in the design parameters to predicted outcomes for the particular engineering problem being studied. The maximum amount of information from each simulation is saved, subject to storage limitations. The creation of the knowledge base is itself a sizeable computational investment, one that pays off if it is used to train a variety of networks for different searches and/or for use in other contexts such as sensitivity analyses. A diagram of the components of the methodology is given in Figure 1. Applications of the methodology have been reported i