Frugal Bayesian optimization for aircraft and HALE drone design

Blackbox simulations are fundamental to modern engineering and industrial applications. In particular, multidisciplinary design optimization of complex systems, such as aircraft, often necessitates high-fidelity solvers. However, they are computationally intensive. To mitigate these prohibitive costs, fast-to-evaluate surrogate models have been adopted across diverse domains, ranging from aerospace engineering and socio-technical transportation systems to meta-modeling in deep learning. Consequently, there is growing interest in Bayesian optimization that leverages Gaussian process surrogates for gaining efficiency, typically concerning mixed-variables and hierarchical design spaces techniques. While they enable the modeling of increasingly complex problems, expensive simulations remain a bottleneck within optimization loops. To address these intractable execution times, parallel computing has become an essential solution. However, large-scale tasks often demand energy-intensive exascale infrastructure, whereas smaller tasks may be more efficiently handled by regional clusters or standard workstations. The objective of this PhD is to evaluate a resource-aware optimization approach that distributes computer experiment workloads across a heterogeneous network of computing nodes to minimize the environmental impact of computations. Given that hierarchical hardware architectures and the availability and consumption of time-dependent power grids vary greatly, the cost and carbon footprint of resource-intensive experiments fluctuate based on hardware configuration, time of day, and location. We model hierarchical spaces to extend thestandard optimization workflow into a system architecture optimization problem with multiple objectives, specifically targeting the environmental impact of computations. The targeted application of this PhD will be the eco-design of high altitude long endurance drone by including the architectural choice of the computational infrastructure in the overall process.