The automatic creation of optimal concepts for mechanical structures in the computer-aided design process has become an important area of research. Continuum topology optimization methods determine the distribution of material within a pre-defined design space and, thus, not only the shape, but also the fundamental geometric layout of a structure. For this task, the majority of the existing, numerical optimization methods requires mathematical gradient information. However, when addressing optimization problems that involve highly non-linear or black-box simulations, it can be difficult to obtain satisfactory results or gradient information at all. In order to provide design concepts also for these types of problems, this thesis presents a generic topology optimization approach. The novel method realizes a self-contained learning component that utilizes physical simulation data to generate a search direction. Based on a continuous problem formulation, every design variable is improved iteratively by a learned update-signal. The individual update-signals are computed from local state features and substitute sensitivities of the design variables. Evolutionary optimization or supervised learning adapt the model parameters for determination of the update-signals to the chosen optimization goal. In empirical studies, the novel method reproduces reference structures with minimum compliance. When applied to a practical problem from the challenging domain of vehicle crashworthiness optimization, specifically the minimization of intrusion, it provides superior design concepts when compared to a frequently applied heuristic method. The results confirm that the proposed method is capable to yield innovative solutions to so far unsolved topology optimization problems.

Fortschr.-Ber. VDI Reihe 20 Nr. 468, ISSN 0178-9473