The proposed architecture instead processes both through two parallel paths. A convolutional path encodes 64×64 images of the airfoil shape, while a fully connected path encodes the Reynolds number and the angle of attack. The extracted features are merged and passed to two regression heads, so that the lift coefficient and the drag coefficient are predicted by separate output branches rather than by a single shared head.

The training data combine UIUC wind-tunnel measurements with results from the XFOIL solver, yielding 12,304 samples. The model was trained in 71 minutes on a standard workstation. The authors present the parameter count and training cost as a central point of the work: a dual-path formulation that couples image-based geometry encoding with flow-parameter conditioning while remaining small enough for training and deployment on commodity hardware.

Accuracy differs markedly between the two outputs. For the lift coefficient, the network achieves a coefficient of determination of 0.940 with a mean absolute percentage error of approximately 0.86 percent. For the drag coefficient, the coefficient of determination is 0.615, a substantially weaker result that the authors attribute to the wider dynamic range and stronger nonlinearity of the drag response, particularly near stall onset.

In conventional workflows that role is filled by computational fluid dynamics, which resolves the flow field numerically at far greater computational expense. The reported 3-millisecond evaluation time per unseen case is the speed advantage the authors highlight against that baseline. The lift branch captures a response that is represented with high fidelity in the training data, while the drag branch faces a harder regression problem within the same tested envelope.

The study's scope is limited to two-dimensional airfoil coefficients over the tested Reynolds number range of 3.92×10⁴ to 5.14×10⁵ and angle-of-attack range of −10.33 to 20 degrees. The drag prediction near stall onset remains the weakest part of the model's performance, and the reported figures describe this specific dataset and architecture rather than a general-purpose aerodynamic solver. As a demonstration, the work shows that image and numerical inputs can be fused in a compact network to predict airfoil aerodynamic coefficients at a speed suitable for rapid evaluation on ordinary computing equipment, with strong lift accuracy, moderate drag accuracy, and a small computational footprint.

Subject of Research: Dual-path CNN for airfoil aerodynamic coefficients prediction using image and numerical data

Article Title: Dual-path CNN for airfoil aerodynamic coefficients prediction using image and numerical data

Article References: Daemiashkezari, M., & Saleh, S. R. (2026). Dual-path CNN for airfoil aerodynamic coefficients prediction using image and numerical data. Aerospace Systems. https://doi.org/10.1007/s42401-026-00561-3

Image Credits: AI Generated

DOI: 10.1007/s42401-026-00561-3

Keywords: Dual-path, airfoil, aerodynamic, coefficients, prediction, image, numerical, data, scientific research