Multi-fidelity Bayesian optimization strategy applied to Overall Drone Design
Rémy Charayron, Thierry Lefèbvre, Nathalie Bartoli, Joseph Morlier · AIAA SCITECH 2023 Forum · 2023
View Video Presentation: https://doi.org/10.2514/6.2023-2366.vid Nowadays, drones can be developed for a wide range of use cases, from infrastructure monitoring to sea rescue, urban mobility or military purposes. Which drone design is best suited for a specific mission? To answer this question, we need to solve a constrained optimization problem based on a multi-disciplinary design model that takes the mission into account. The model generally being a computationally expensive numerical model whose gradients are not available all the time encourages us to consider a Bayesian optimization approach. Such strategy is well known to achieve a trade-off between exploitation and exploration in order to find interesting minimal area with a reduced number of function evaluations. A multi-fidelity approach can improve even more the computational efficiency of the Bayesian optimization strategy. In this work, we aim at designing a fixed-wing drone (fully electric) for long range surveillance mission. Two fidelity level electric drone models are developed. For a given mission requirement, the final battery state of charge is optimized with respect to drone design variables. Optimizations are performed on several missions using both a mono and a multi-fidelity Bayesian optimization strategy. The interest of using a multi-fidelity method for overall drone design has been assessed. The multi-fidelity super-efficient global optimization algorithm (MFSEGO) appeared to need less budget to reach convergence than the mono-fidelity algorithm and to be more robust to the initial design of experiments.