A hybrid machine learning framework for rotor performance prediction

Recent advancements in electric propulsion technology have created significant demand for rotor performance models that can be used for preliminary design and optimization of new class of electric vertical takeoff and landing aircraft. Existing high-fidelity computational fluid dynamics models, while accurate, are computationally expensive and too slow for real-time and optimization applications. This paper presents a novel approach, NN-BET, which combines neural networks (NN) with blade element theory (BET) to predict rotor loads based on rotor geometrical features, operational parameters, and airfoil characteristics. This hybrid surrogate model employs a neural network to estimate an intermediate aerodynamic parameter (mean inflow), which is then integrated into the BET framework to calculate aerodynamic loads. Incorporation of NN architecture inside a blade-element framework makes it an order of magnitude faster than a traditional blade element momentum theory model, one of the fastest physics-based models, while maintaining similar accuracy. To assess the influence of integrated physical principles, the performance of the developed NN-BET model is compared to that of a conventional neural network model. Both models are trained on the same dataset with identical neural network architectures and hyperparameters to maintain a fair and consistent comparison. The results show that the NN-BET model surpasses the conventional NN model in terms of predictive accuracy. Given the rapid computational speed and reliable accuracy of the developed machine learning-based model, it is well suited for optimization and design applications, as well as multidisciplinary frameworks requiring efficient rotor dynamics modeling.