Three-dimensional aerodynamic design optimization using discrete sensitivity analysis and parallel computing

Abstract

A hybrid automatic differentiation/incremental iterative method was implemented in the general purpose advanced computational fluid dynamics code (CFL3D Version 4.1) to yield a new code (CFL3D.ADII) that is capable of computing consistently discrete first order sensitivity derivatives for complex geometries. With the exception of unsteady problems, the new code retains all the useful features and capabilities of the original CFL3D flow analysis code. The superiority of the new code over a carefully applied method of finite-differences is demonstrated. A coarse grain, scalable, distributed-memory, parallel version of CFL3D.ADII was developed based on stripmining. In this data-parallel approach, an identical copy of CFL3D.ADII is executed on each processor with different derivative input files. The effect of communication overhead on the overall parallel computational efficiency is negligible. However, the fraction of CFL3D.ADII duplicated on all processors has significant impact on the computational efficiency. To reduce the large execution time associated with the sequential 1-D line search in gradient-based aerodynamic optimization, an alternative parallel approach was developed. The execution time of the new approach was reduced effectively to that of one flow analysis, regardless of the number of function evaluations in the 1-D search. The new approach was found to yield design results that are essentially identical to those obtained from the traditional sequential approach but at much smaller execution time. The parallel CFL3D.ADII and the parallel 1-D line search are demonstrated in shape improvement studies of a realistic High Speed Civil Transport (HSCT) wing/body configuration represented by over 100 design variables and 200,000 grid points in inviscid supersonic flow on the 16 node IBM SP2 parallel computer at the Numerical Aerospace Simulation (NAS) facility, NASA Ames Research Center. In addition to making the handling of such a large problem possible, the use of parallel computation provided significantly reduced overall execution time and turnaround time.