@phdthesis {pub4523,
	title = {Level Set Topology Optimization for Crashworthiness using Evolutionary Algorithms and Machine Learning},
	author = {Mariusz Bujny},
	year = {2020},
	abstract = {Due to the rising complexity of the development process in the automotive industry, it becomes very difficult to rely solely on the engineering intuition when designing car components. At the same time, the use of simulation methods in industry is now a common standard, leading to a transformation of the traditional design process towards the model-based concept. As a consequence, automatic generation of mechanical structures based on optimization algorithms is more and more frequently used in practice. In particular, formulation of a design task as a topology optimization problem gives the computer program the most flexibility, allowing for a redistribution of the material within a given design space, to maximize performance metrics for the defined loading conditions. Most of the commonly used topology optimization methods utilize analytical sensitivity information to perform efficient gradient-based search even for problems involving millions of design variables. However, some important optimization problems, such as the ones in structural crashworthiness, exhibit very high complexity, reflected in strong nonlinearity, high levels of numerical noise, bifurcations, and discontinuities of the considered objectives and constraints. Hence, the gradient-based methods are usually not directly applicable in such cases and alternative approaches, based on strong simplifications in the modeling or heuristic assumptions, are used instead. As a result, they are suitable mainly for specific cases, and a further research to develop more general methods is needed. Therefore, this thesis proposes a topology optimization approach based on evolutionary algorithms and a low-dimensional level-set representation, allowing for an optimization of arbitrary quantifiable criteria using high-fidelity explicit crash simulations. The method is thoroughly validated based on the standard linear elastic benchmark problems and compared to the state-of-the-art crash topology optimization methods using academic test cases as well as a real-world optimization problem. The numerical experiments show that considerably better structures can be obtained with the proposed method, however, at the cost of a high number of necessary crash simulations. To mitigate that, this thesis proposes to use efficiently the available information by integrating machine learning at different levels of the optimization process: for modeling of the responses from the expensive high-fidelity simulations, for approximating sensitivity information used in a hybrid gradient-enhanced evolutionary approach, and for prediction of favorable topological variations within an adaptive topology optimization. The results show that significant performance improvements can be obtained by incorporating machine learning techniques into the evolutionary search. Finally, the generic character of the proposed methods allows potentially for addressing a wide spectrum of non-standard structural optimization tasks, including problems in manufacturing, soft robotics, compliant mechanism design, and many others.},
	publisher = {Technical University of Munich},
	booktitle = {Technical University of Munich}
}