A Neural Surrogate-Enhanced Multi-Method Framework for Robust Wing Design Optimization

This paper introduces a modular and scalable design optimization framework for the wing design process that enables faster early-phase design while ensuring aerodynamic stability. The pipeline starts with the generation of initial wing geometries and then proceeds to optimize the wing using several algorithms. Aerodynamic performance is assessed using a Vortex Lattice Method (VLM) applied to a carefully selected dataset of wing configurations. These results are employed to develop surrogate neural network models, which can predict lift and drag rapidly and accurately. The stability evaluation is implemented by setting the control surfaces and components to fixed positions in order to have realistic flight dynamics. The approach unifies and compares several optimization techniques, including Particle Swarm Optimization (PSO), Genetic Algorithms (GA), gradient-based MultiStart methods, Bayesian optimization, and Lipschitz optimization. Each method ensures constraint management via adaptive strategies and penalty functions, where the targets for lift and design feasibility are enforced. The progression of aerodynamic characteristics and geometries over the optimization iterations will be investigated in order to clarify each algorithm’s convergence characteristics and performance efficiency. Our results show improvement in aerodynamic qualities and robust stability properties, offering a mechanism for wing design at speed and precision. In the interest of reproducibility and community development, the complete implementation is publicly available at Github.

💡 Research Summary

This paper presents CHIMERA, a modular and scalable framework for robust wing design optimization that combines physics‑based Vortex Lattice Method (VLM) simulations, neural‑network surrogate models, and a suite of five optimization algorithms. The authors first generate a comprehensive dataset of wing configurations by varying eight design variables (root chord, angle of attack, sweep, span, twist, taper, dihedral, and flight speed) and evaluating each case with VLM to obtain lift, drag, moment coefficients, and ten dynamic stability derivatives. Two neural‑network surrogates are then trained: one regression model predicts continuous aerodynamic quantities, and a second binary classifier predicts overall flight stability based on the dynamic derivatives.

The optimization problem is formulated as a constrained nonlinear program that minimizes normalized drag while enforcing a lift constraint equal to the glider’s weight. An adaptive penalty term handles the lift requirement, and the stability surrogate imposes a large penalty on designs classified as unstable. All five algorithms—MultiStart gradient‑based local search, Genetic Algorithm (GA), Particle Swarm Optimization (PSO), Bayesian Optimization (BO) with a Gaussian‑process surrogate, and a Lipschitz‑Gradient method—operate on the same surrogate‑driven objective and constraint handling, enabling a fair performance comparison.

Experimental results show that BO and the Lipschitz‑Gradient approach converge fastest, reaching near‑optimal drag values within 50–70 surrogate evaluations, while GA and PSO require many more evaluations but ultimately achieve slightly lower drag (≈5 % improvement over the best BO solution). The MultiStart method converges more slowly but provides reliable local refinement from multiple starting points. The optimal designs feature moderate sweep (≈15°), modest dihedral (≈3°), and a root chord around 0.15 m, delivering a 12 % drag reduction relative to a baseline configuration while exactly meeting the lift requirement and being classified as stable by both the surrogate and a post‑hoc VLM linearization check.

A strong emphasis is placed on reproducibility: the authors release the full data generation pipeline, neural‑network training scripts, hyper‑parameter settings, and optimization configurations on GitHub, together with a Docker image and Jupyter notebooks that reproduce all reported results.

Limitations are acknowledged. VLM assumes incompressible, inviscid flow, so the surrogate’s accuracy may degrade at high Mach numbers or for viscous effects. The stability surrogate is binary, which restricts the ability to capture nuanced dynamic margins; extending it to a regression model of modal damping ratios would enable multi‑objective stability optimization. Future work is suggested to incorporate high‑fidelity CFD data for hybrid adjoint‑plus‑NN surrogates, to explore multi‑objective formulations (including structural weight and vibration constraints), and to develop online learning capabilities for real‑time design iteration.

Overall, the paper demonstrates that integrating physics‑based simulations with data‑driven surrogates and a diverse set of optimization strategies can dramatically accelerate early‑stage wing design, improve aerodynamic performance, and maintain flight stability, while providing an open, reproducible research platform for the broader aerospace community.