Tuning Machine Learning Outcomes for Airfoil Aerodynamics Seed Sensitivity in Predictive Performance
Abstract. Accurate aerodynamic performance prediction remains critical in preliminary design and optimization workflows. This study proposes a hybrid deep learning framework that combines convolutional neural networks (CNNs), operating directly on raw airfoil geometries, with two fully connected branches that process engineered shape descriptors and flow parameters such as angle of attack and Reynolds number. This model is trained on a high-resolution dataset generated via XFOIL simulations, covering 91 NACA 4-digit airfoils across 200 aerodynamic conditions each, spanning Reynolds numbers from 500,000 up to 5,000,000 and angles of attack between -5˚ and 14˚. A key novelty introduced in this work is the investigation of how random seed initialization influences predictive accuracy. In machine learning, a random seed is a fixed numerical input that initializes the pseudo-random number generator responsible for operations such as weight-initialization and data shuffling. While typically treated as a background setting, the seed plays a critical role in defining the starting conditions of model training and can directly influence final performance outcomes. Results show that models with identical architecture and data can differ in performance by up to 250% depending solely on the seed used during training. To mitigate this variability, ensemble strategies based on seed diversity were also explored, leading to greater consistency and generalization. The top-performing single-seed model tested (seed 0) achieved a mean absolute percentage error (MAPE) of 1.1% for lift coefficient and 0.57% for drag coefficient, with R2 values of 0.9998 and 0.9954, respectively. The best ensemble built from three diverse seeds (C2) achieved slightly higher generalization: 1.43% MAPE CL, 1.19% MAPE CD, with R2 values of 0.9999 and 0.9968. All models tested, spanning ten seeds and five ensembles, achieved R2 scores exceeding 0.97. The results emphasize that treating seed selection as a fixed default can overlook significant variability, and that ensemble calibration can serve as a robust strategy to improve prediction reliability.