Clustering-Enhanced MLP for Aerodynamic Surface Pressure Prediction
Please login to view abstract download link
The analysis of aerodynamic forces on lifting surfaces is a fundamental task in the aerospace industry and is traditionally addressed using CFD simulations and empirical wind tunnel–based models. Although these approaches provide high-fidelity results, their computational cost limits their applicability for rapid evaluations and preliminary design stages. Machine learning models offer an attractive alternative by significantly reducing computational effort while maintaining acceptable accuracy. Among supervised learning techniques, the Multilayer Perceptron (MLP) is widely used to model nonlinear problems based on reference datasets. However, certain aerodynamic phenomena exhibit strong nonlinear and localized effects, such as flow field discontinuities at transonic and sonic speeds, which remain challenging to capture accurately. Previous studies have explored learning-based approaches for these regimes with promising results. This work aims to improve predictive performance by decomposing the problem into smaller regions using a Gaussian Mixture Model (GMM) clustering algorithm. The flow field is first classified according to selected variables and subsequently evaluated using region-specific regression models tailored to each cluster. The proposed approach achieves performance improvements of up to 50% compared to a standard MLP, enabling a more accurate representation of nonlinear effects and complex aerodynamic phenomena.
