Creating surrogate models for an air and missile defense simulation using design of experiments and neural networks

Abstract

This paper demonstrates a method of constructing multiple linked surrogate models of a high-fidelity air and missile defense simulation using design of experiments to generate labeled data for neural network models. The surrogate models are used to predict the number of incoming missiles destroyed and the number of interceptors launched from a multi-layered defense composed of three different air defense systems intercepting both ballistic and cruise missiles without the need for time intensive simulation runs. A single model that predicts all outcomes was first attempted, but was shown to have inadequate prediction capabilities. The working setup uses multiple surrogate models that are linked to allow information to pass between each model. The paper demonstrates how to develop the surrogate models using a notional example, and how to link these surrogate models together using time to impact for the missiles. The same methodology also allows the same surrogate model to switch between ballistic and cruise missile engagements. When run on a desktop computer, a 30 Monte Carlo set of the notional example took several minutes to complete; however, this proof of principal implementation of the surrogate models was able to predict the mean number missiles destroyed or the mean number of interceptors fired to within one missile nearly instantaneously.