Aerodynamic design optimization is an important problem in aircraft design that depends on the interplay between a numerical optimizer and a high-fidelity flow physics solver. Derivative-based, first and (quasi) second order, optimization techniques are the de facto choice, particularly given the availability of the adjoint method and its ability to efficiently compute gradients at the cost of just one solution of the forward problem. However, implementation of the adjoint method requires careful mathematical treatment, and its sensitivity to changes in mesh quality limits widespread applicability. Derivative-free approaches are often overlooked for large scale optimization, citing their lack of scalability in higher dimensions and/or the lack of practical interest in globally optimal solutions that they often target. However, breaking free from an adjoint solver can be paradigm-shifting in broadening the applicability of aerodynamic design optimization. We provide a systematic benchmarking of a select sample of widely used derivative-based and derivative-free optimization algorithms on the design optimization of three canonical aerodynamic bodies, namely, the NACA0012 and RAE2822 airfoils, and the ONERAM6 wing. Our results demonstrate that derivative-free methods are competitive with derivative-based methods, while outperforming them consistently in the high-dimensional setting. These findings highlight the practical competitiveness of modern derivative-free strategies, offering a scalable and robust alternative for aerodynamic design optimization when adjoint-based gradients are unavailable or unreliable.
Our choice of derivative-based and derivative-free algorithms for benchmarking is by no means exhaustive. However, we select algorithms that are widely found in well-established open-source optimization libraries, and we ensure there is sufficient diversity in their respective methodologies.
Our study is not an evaluation of global versus local optimization approaches to ASO. Rather, we are interested in the practical utility of derivative-free approaches to ASO. In this regard, we entertain both local and global derivative-free approaches in this study.
As a first step, we restrict ourselves to moderate dimensions (< 50); higher dimensional investigation is reserved for future work.
Note that we don’t restart the local optimization algorithms with a randomized choice of \mathbf{x}_0. Instead, \mathbf{x}_0 is chosen from the initial pool of seed points provided to BO that corresponds to the best feasible objective value. blackThis way, we ensure we don’t disadvantage local optimization methods with a starting point worse than that of BO. On the other hand, BO is repeated 3 times particularly to address the stochasticity of the seed point selection.
blackWe do have strong reasons for restricting our study to SciPy, as opposed to other libraries used for aerodynamic design optimization such as SNOPT [gill2005snopt]. First, SciPy is a widely adopted research and industry standard software (see, e.g., [virtanen2020scipy]). Industry standard machine learning libraries still use SciPy routines under the hood (see, e.g., [balandat2020botorch]). Second, SNOPT has restrictions in its usage – for academic research and student learning, an unrestricted version is available but with a time limit (https://ccom.ucsd.edu/~optimizers/downloads/). SciPy, on the other hand, is openly available and is actively under maintenance and development. Finally, SNOPT presents a sparse version of an SQP algorithm with a limited-memory approximation to the Hessian of the Lagrangian – the software is suitable for smooth nonconvex programs. The same problem class is also targeted by SciPy’s SLSQP and by SciPy’s Trust-constrained (a trust-region constrained method that, for equality-constrained problems, implements a Byrd–Omojokun trust-region SQP [omojokun1989trust] variant and, with inequalities, switches to a trust-region interior-point approach). In other words, our SciPy baseline already exercises the same family of methods as SNOPT. For these reasons, we choose SciPy for this work that also promotes replicability of results.
The text on this page is from arXiv:2505.09088v2 by Punya Plaban, Peter Bachman and Ashwin Renganathan, under the licence CC BY 4.0. The corpus made these changes: it read the items out of the LaTeX source, it expanded the macros of the paper, it put each proof after its statement, and it converted the text to HTML. The formulas are the TeX of the paper, and KaTeX renders them. A formula or a reference can look different from the PDF. The page is published under CC BY 4.0.
The record of each item is in analysis/objects/2505/2505.09088.jsonl.