Sequential Quadratic Programming (SQP) for optimal control in direct numerical simulation of turbulent flow

详细信息    查看全文

• 作者：Hassan Badreddine ; Stefan V ; ewalle ; Johan Meyers
• 关键词：Sequential Quadratic Programming ; Damped limited-memory BFGS ; Turbulent mixing layer ; Optimal control ; Direct Numerical Simulations ; Adjoint equations
• 刊名：Journal of Computational Physics
• 出版年：2014
• 出版时间：1 January, 2014
• 年：2014
• 卷：256
• 期：Complete
• 页码：1-16
• 全文大小：1768 K

文摘

The current work focuses on the development and application of an efficient algorithm for optimization of three-dimensional turbulent flows, simulated using Direct Numerical Simulation (DNS) or Large-Eddy Simulations, and further characterized by large-dimensional optimization-parameter spaces. The optimization algorithm is based on Sequential Quadratic Programming (SQP) in combination with a damped formulation of the limited-memory BFGS method. The latter is suitable for solving large-scale constrained optimization problems whose Hessian matrices cannot be computed and stored at a reasonable cost. We combine the algorithm with a line-search merit function based on an -norm to enforce the convergence from any remote point. It is first shown that the proposed form of the damped L-BFGS algorithm is suitable for solving equality constrained Rosenbrock type functions. Then, we apply the algorithm to an optimal-control test problem that consists of finding the optimal initial perturbations to a turbulent temporal mixing layer such that mixing is improved at the end of a simulation time horizon T. The controls are further subject to a non-linear equality constraint on the total control energy. DNSs are used to resolve all turbulent scales of motion, and a continuous adjoint formulation is employed to calculate the gradient of the cost functionals. We compare the convergence speed of the SQP L-BFGS algorithm to a conventional non-linear conjugate-gradient method (i.e. the current standard in DNS-based optimal control), and find that the SQP algorithm is more than an order of magnitude faster than the conjugate-gradient method.