A Hybrid Aerodynamic-Interference-Aware Model for Multi-Rotor Platforms

Traditional aerodynamic models often struggle to accurately capture complex aerodynamic interference, particularly in non-collinear rotor configurations where cross-rotor interactions significantly affect rotor forces and torques. While advanced models using Computational Fluid Dynamics (CFD) can model these complexities, their computational intensity limits their usability in real-time simulation and control applications. Conversely, data-driven models such as Artificial Neural Networks (ANN) provide real-time predictive capabilities but struggle with issues related to sample efficiency, generalization and physical consistency.

This thesis presents a novel hybrid aerodynamic-interference-aware model that combines interference-unaware physics-based modeling with deep learning to accurately predict the aerodynamic forces and torques acting on the rotor of the OmniMorph UAV. Multiple neural network architectures, including Multilayer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), Long Short-Term Memory Networks (LSTMs), and Temporal Convolutional Networks (TCNs), were integrated both as end-to-end and hybrid models and systematically compared. 

Experimental results demonstrate that the proposed hybrid models significantly outperform both traditional and state-of-the-art methods in predictive accuracy. The Hybrid Temporal Convolutional Network, in particular, achieves up to a 97.7\% reduction in mean percentage error compared to traditional models. Its inference time allows processing data over 250 times quicker than the data sampling rate, making it highly suitable for real-time applications. Our comprehensive evaluation provides insights into the strengths and limitations of different models, guiding future neural network-based design in aerodynamic modelling.

Lastly, we provide some recommendations for future researchers, including integrating the hybrid model with real-time control strategies, such as Model Predictive Control, and implementing an online-offline training schema to improve the model's adaptability and robustness in dynamic flight conditions.