Investigation of Adjoint Based Shape Optimization Techniques in NASCART-GT using Automatic Reverse Differentiation

Abstract

Automated shape optimization involves making suitable modifications to a geometry that can lead to significant improvements in aerodynamic performance. Currently available mid-fdelity Aerodynamic Optimizers cannot be utilized in the late stages of the design process for performing minor, but consequential, tweaks in geometry. Automated shape optimization involves making suitable modifications to a geometry that can lead to significant improvements in aerodynamic performance. Currently available mid-fidelity Aerodynamic Optimizers cannot be utilized in the late stages of the design process for performing minor, but consequential, tweaks in geometry. High-fidelity shape optimization techniques are explored which, even though computationally demanding, are invaluable since they can account for realistic effects like turbulence and viscocity. The high computational costs associated with the optimization have been avoided by using an indirect optimization approach, which was used to dcouple the effect of the flow field variables on the gradients involved. The main challenge while performing the optimization was to maintain low sensitivity to the number of input design variables. This necessitated the use of Reverse Automatic differentiation tools to generate the gradient. All efforts have been made to keep computational costs to a minimum, thereby enabling hi-fidelity optimization to be used even in the initial design stages. A preliminary roadmap has been laid out for an initial implementation of optimization algorithms using the adjoint approach, into the high fidelity CFD code NASCART-GT.High-fidelity shape optimization techniques are explored which, even though computationally demanding, are invaluable since they can account for realistic effects like turbulence and viscocity. The high computational costs associated with the optimization have been avoided by using an indirect optimization approach, which was used to dcouple the effect of the flow field variables on the gradients involved. The main challenge while performing the optimization was to maintain low sensitivity to the number of input design variables. This necessitated the use of Reverse Automatic differentiation tools to generate the gradient. All efforts have been made to keep computational costs to a minimum, thereby enabling hi-fidelity optimization to be used even in the initial design stages. A preliminary roadmap has been laid out for an initial implementation of optimization algorithms using the adjoint approach, into the high fidelity CFD code NASCART-GT.