Structural optimization for fail-safe designs by machine learning

• Applicant: Professor Dr.-Ing. Benedikt Kriegesmann
• Subject Area: Lightweight Construction, Textile Technology ; Mechanics
• Term: since 2023
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 515743381

In certain industries like the aerospace industry safety-critical components have to be designed so that they are damage tolerant. One way to ensure damage tolerance is a fail-safe design which is able to sustain the applied load even if one structural member/load path fails. The requirement for fail-safety can be embedded into design optimization as a constraint. However, the computational cost of solving such an optimization problem is too expensive to apply this approach to problems of industrial relevance. Especially in the context of topology optimization, which provides the highest design freedom, the number of simulations that need to be run in order to account for all damage scenarios is extremely high. The structures which result from such topology optimizations have similar global load paths as structures that are optimized without consideration of fail-safety. However, the fail-safe designs have multiple struts along the load paths and more connecting elements. One fast and practicable approach is to run an optimization without consideration of fail-safety to determine the main load paths. Then, based on the experience of the design engineer, a design with redundant load paths is derived. This design process is obviously dependent on the experience of the designer. This is a problem especially for topology optimization due to the large amount of possible solutions. The objective of the current proposal is to replace the engineering experience by an artificial neural network. For that, a large number of combinations of different design spaces, boundary conditions and loads is considered. For each combination, a “plain” topology optimization as well as a topology optimization considering fail-safety is performed. The obtained topologies serve as data for training an artificial neural network. When designing a new structure, only a topology optimization without consideration of fail-safety shall be conducted. The obtained topology is the input for the already trained artificial neural network, which then suggests a redundant topology. Afterwards, this topology is fine-tuned with a fast shape optimization approach. While existing approaches for topology optimization considering fail-safety regard the compliance as objective function, the current proposal also addresses eigenfrequencies. Here, the classical maximization of the minimum eigenfrequency is considered as objective as well as minimizing the change of eigenfrequencies in case of partial failure.

• DFG Programme: Research Grants