Data-efficient AI models for aerodynamic coefficient prediction: a case study on NACA airfoils with minimal CFD runs

Abstract

Abstract Background High-fidelity computational fluid dynamics (CFD) remains a bottleneck during early aerodynamic design, where many candidate configurations must be screened under strict computational budgets. Data-efficient artificial intelligence (AI) surrogates offer a promising solution to recover aerodynamic coefficients with minimal simulation effort. Within this context, achieving comparable predictive performance using a fraction of the CFD data is essential for efficient surrogate-based design exploration. To reflect real aerodynamic design constraints, this study explicitly investigates how far CFD-generated datasets can be reduced while maintaining predictive reliability, establishing a quantitative balance between accuracy and computational cost. Methods A validated dataset was generated using Reynolds-averaged Navier–Stokes (RANS) simulations with the Spalart–Allmaras turbulence model for four NACA airfoil families, covering a range of Reynolds numbers and angles of attack. To emulate resource-constrained scenarios, the dataset was systematically reduced to 30%, 25%, 20%, 15%, 10%, and 5% of its original size. Ten widely used machine learning algorithms were benchmarked against feedforward backpropagation neural networks with varying hidden layer sizes. Results Across the tested configurations, models maintained strong predictive performance even with substantially reduced datasets. At a 10% training ratio, some models achieved R 2 values up to 0.98398 with low mean absolute error, while simulation time decreased from approximately 36 h 40 min to about 3 h 45 min. However, reducing the dataset to 5% resulted in a measurable decline in accuracy, particularly for lift predictions and for deeper neural networks. Conclusions This study demonstrates that AI surrogates can be integrated into CFD workflows to significantly reduce computational cost while preserving predictive accuracy. The findings establish a practical framework for dataset reduction, paving the way for extending this methodology to broader aerodynamic configurations in future studies.