Efficiency of an Adjoint Industrial CFD Code

Sensitivity analysis with the aim of design optimization is a growing area of interest in Computational Fluid Dynamics (CFD) simulations. However, one of the major challenges is to deal with a large number of design variables for largescale industrial applications. One of the effective solution approaches is to compute adjoint-based sensitivities in the differentiated CFD code. In this paper, we develop a discrete adjoint code for an unstructured pressure-based steady NavierStokes solver using Algorithmic Differentiation (AD) by operator overloading (OO) tool. To reduce the huge memory requirement of the adjoint code we apply effective techniques by implementation of checkpointing schemes and by symbolic differentiation of the iterative linear solver. We combine the flexibility of an operator overloading tool with the efficiency of an adjoint code generated by source transformation through coupling these approaches. Moreover, we improve the performance of the adjoint computation by exploiting the mathematical aspects of the involved fixed-point iteration through reverse accumulation. We compare the effectiveness of these methods in terms of reduction of the numerical cost and accuracy of the sensitivities for the optimization of a vehicle climate duct industrial test case.