Neural Network Based Aerodynamic Loads Estimation for Launch Vehicle Preliminary Design

This paper presents a methodology for estimating aerodynamic loads in supersonic regimes for the preliminary design of a launch vehicle using a long-short term memory model, with potential applications in estimating internal structural loads for structural sizing purposes. The initial stage of the methodology is the generation of an aerodynamic dataset based on the programmatic generation of a computational fluid dynamics mesh. This presents the launch vehicle profile and flow properties as inputs and the resulting lateral loads distribution as outputs. The subsequent stage is the training of a space convolution model based on a neural network on the aforementioned dataset. Finally, the trained model is tested with additional data, and the results are compared with computational fluid dynamics simulations.