Synergistic Computing: Combining CFD and Neural Networks for Maneuvering Simulation

Abstract

A method has been developed that enables the synergistic computing of vehicle maneuvering simulations based on a combination of computational fluid dynamics (CFD) and feedforward neural networks (NNs). The CFD code is used to provide a physics based solution applicable to both design as well as maneuvering simulation. As would be expected, the CFD solutions are not perfect and there is a finite level of error in the predicted total forces and moments. The NNs are developed to adaptively correct, add, the expected residual error in the total forces and moments. Estimates of the residual forces and moments can be obtained by comparing the predictions produced by the CFD simulation and maneuvering data at both model and fullscale. Three NNs are used to predict the moment residuals, correction of the force residuals proved to be unnecessary. A lesioning, pruning, method for NNs was extended in order to determine a suitable interface between the inputs and outputs (I/O) of both codes. This synergistic interface is one of the critical steps required in combining these two codes to produce an accurate simulation tool. Simulation results are shown comparing the predicted maneuvers for the combined codes to those of the stand alone CFD code. The resultant simulation predictions are shown to be far more accurate than either of the two separate simulation approaches, CFD or RNN, when used individually. In 100 percent of the cases, the synergistic code is more accurate. The results indicate that it is possible, within limits, to create an essentially “perfect” simulation of the vehicle. Perfect, in this case, being no real difference between the simulation and experiment on the majority of the maneuvers and only minor differences on the very worst cases. These tools have also been extended to other fast attack class submarines with comparable results. These solutions demonstrate the opportunity for developing improved physics based simulation tools, for highly nonlinear vehicle motions, based on combining multiple different codes into a single tool. In these cases, this new class of simulation tools can provide simulations with greatly improved accuracy.