Approximating a finite element model by neural network prediction for facility optimization in groundwater engineering

A common technique for modeling complex systems in many computational engineering domains is the finite element method. Analyzing these systems using the finite element method involves the application of time consuming methods of numerical simulation. Optimizing these systems often implies many hundreds of these analyses leading to an excessive computational load and therefore an impractical runtime of the optimization process. In this paper, an approach is presented to utilize predictions computed by a neural network to approximate the results of a simulation system based on the finite element method, yielding a precision of the approximation of at least 99%. This approximation is integrated into an optimization strategy. The prediction-based optimization approach is demonstrated by solving a facility optimization problem in groundwater engineering. The results demonstrate that the computation time can be reduced by at least 60%.