Designing future complex systems requires the integration of disruptive technologies and weighting conflicting stakeholder needs. The design of the system architecture, however, usually feature a very large design space and require the application of expensive, physics-based simulation due to a lack of historical data. System Architecture Optimization (SAO) can partially solve this conflict by applying numerical optimization algorithms to the system architecture design process [1].

The state-of-the-art algorithms capable of solving such problems are Bayesian Optimization (BO) algorithms: optimization algorithms that use surrogate models, usually Gaussian Process (GP) models, to learn the behavior of the design space and try to predict where the optimal design points lie using an infill criterion. Research shows that they can tackle relatively simple problems, however still have some limitations in terms of number of design variables and number of optimization objectives [1].

The goal of this student project is to improve the performance of optimization algorithms for SAO, by (partially) lifting aforementioned limitations, and by improving the predictive capabilities of the surrogate models. In particular, because the system architecture is modeled using directed graphs, two techniques may be interesting:

1. Graph kernels [2]: GP kernels that are tuned directly on the graph structure.

2. Graph neural networks [3]: Neural networks trained on the graph structure and data itself.

The project can be completed as a master’s internship (6 months) or thesis. Depending on the quality of the work, a conference paper might result from it.

Tasks

1. • Improve the performance of existing optimization algorithms with respect to:

     • The number of design variables (enable the algorithms to support up to approx. 100 design variables)

     • The number of objectives (enable the algorithms to support approx. 10 objectives)

   • Investigate new techniques for improving surrogate model prediction:

     • Graph kernels for Gaussian Processes

     • Graph neural networks

   • Investigating mutual interactions between existing and newly implemented techniques

   • Compare optimization algorithm performance against existing optimization algorithms

   • Apply to realistic system architecting problems

Student Profile

For this project, we are looking for a master’s student, with a passion for and experience with: numerical optimization and programming (Python). Knowledge of aircraft design and/or systems engineering helps. You are willing to dive deep into a topic, while also keeping the bigger picture in view as is needed for the integration of different engineering disciplines. You are able to work independently and be able to communicate problems and results effectively.

Logistics

You will work in the MDO/MBSE group of the Institute of System Architecture in Aeronautics in Hamburg, Germany. The DLR is the national aeronautics and space research center of Germany, with over 10k employees distributed over 52 institutes. The DLR offers a stimulating environment of passionate researchers working on all topics in aerospace engineering. More information: https://www.dlr.de/en/careers/your-entry/your-career-level/students-doctoral-candidates.

At the institute in Hamburg, topics include overall aircraft design, aircraft design methods (MDO, MBSE, KBE), and cabin systems design. We work in many national and international projects with partners from all over the world. The working language is English and knowledge of German is not required.

You will receive a student researcher contract from the DLR and live and work in Hamburg during the time of your contract.

For more information, please contact: Jasper Bussemaker

Literature

1. Jasper H. Bussemaker. System Architecture Optimization: Function-Based Modeling, Optimization Algorithms, and Multidisciplinary Evaluation. PhD dissertation, TU Delft, jul 2025.

2. G. Nikolentzos, G. Siglidis, and M. Vazirgiannis. “Graph Kernels: A Survey”. In: Journal of Artificial Intelligence Research 72 (Nov. 2021), pp. 943–1027. DOI: 10.1613/jair.1.13225

3. Z. Wu et al. “A Comprehensive Survey on Graph Neural Networks”. In: IEEE Transactions on Neural Networks and Learning Systems 32.1 (Jan. 2021), pp. 4–24. DOI: 10.1109/tnnls.2020.2978386