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

TL;DR

This study introduces CHIMERA, a modular framework that combines vortex-lattice data, dual neural surrogates (aerodynamic and stability), and a multi-method optimization ensemble to accelerate robust wing design for ultralight gliders. By evaluating five optimization strategies within a unified surrogate-based pipeline, the work demonstrates both speed and convergence characteristics across multimodal design spaces, while highlighting extrapolation challenges and the need for stability-aware objectives. The results show credible drag reductions and feasible lift compliance across methods, with trade-offs in stability classification and surrogate fidelity that motivate future multi-fidelity and physics-augmented enhancements. The authors also provide a public code release to support reproducibility and further development in surrogate-assisted aero-optimization.

Abstract

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.

Figures (16)

• Figure 1: Box plots of the input design variables showing coverage and distribution within the prescribed bounds.
• Figure 2: Workflow of dataset generation, preprocessing, and surrogate model training.
• Figure 3: Airfoil geometry (NACA 2412) used across the span.
• Figure 4: Architecture of both neural networks (White: Input layer, Dark Grey: Fully connected layer, Red: ReLU activation, Light Grey: Add layer, Black: Output layer). Residual connections exist between hidden layers. (2.4 million trainable parameters).
• Figure 5: Aerodynamics Neural Network process.