A Fast, Memory-Efficient Panel Method for Large-Scale Multidisciplinary Design Optimization Under Uncertainty Using Graph-Based Modeling

Comprehensive large-scale multidisciplinary design optimization for aircraft design has become feasible thanks to recent advances in automatic adjoint-based sensitivity analysis. However, a major limitation of existing methods and previous demonstrations is the use of low-fidelity models to address key disciplines. Aerodynamics, one of these key disciplines, is currently restricted to very low or high fidelity methods. Panel methods bridge this gap in modeling fidelity; however, existing tools are not suited for large-scale MDO because they lack adjoint-based sensitivity analysis needed for efficient gradient computation. In addition, panel methods naturally incur a prohibitive ��2 cost. In this paper, we address these needs by presenting a modern, graph-based panel method with automatic adjoint-based sensitivity analysis. We discuss novel methods focused on reducing the prohibitive quadratic cost with respect to grid size by using a new approach we call partitioned vectorization, and present a method incorporating POD that reduces the memory cost for solving the linear system to �� (��). With this, we present results on numerical experiments comparing timing and memory for other standard approaches, showing memory improvements by up to two orders of magnitude with minimal increase in computation time. We then present a large-scale multidisciplinary design optimization under uncertainty (MDOUU) demonstration on the blended wing body aircraft concept using our graph-based panel method. We highlight the importance of robustness to input variation by identifying significant constraint violations of the MDO design in the MDOUU setting.