| The topology design of structures has a major impact on their cost and performance. However, currently available methods for topology optimization cannot be effectively applied to large-scale structures. This research developed two new methods for this purpose.; The first method, called hierarchical, interactive, and metamodel-based optimization (HIMO), combines a sizing optimizer with a metamodeling technique. At the lower level, for each candidate topology design, a sizing optimizer finds feasible and optimal solutions in terms of sizing variables (plate thickness in continuum structures). All performance constraints such as stress, displacement, stability, etc., are handled only at this level. At the upper level, a metamodel is built to fit all the optimal solutions found at the lower level. The metamodel is then used to find the optimal topology design. Only the objective function (e.g. weight) is approximated. The number of topology design variables is much smaller than required by other topology optimization methods, thus large-scale structural systems can be optimized. HIMO was applied to two design projects, resulting in 18% and 36% weight savings, and significant reductions in manufacturing cost.; The second method, called sizing optimizer for topology optimization (SOTO), directly uses a sizing optimizer to optimize topology as well as thickness, using finite element analysis as the simulation tool. The thinner elements are gradually deleted to achieve improved or even optimal topology. The design problem is then reformulated with much fewer design variables for final sizing optimization. For complex 3D problems, an associated partial ground structure approach is also proposed. SOTO was applied to several numerical test problems, and then to a real design project, demonstrating the applicability and efficiency of SOTO.; In order to improve metamodeling used in HIMO, through more effective sampling, 18 experimental design methods were compared and some fundamental issues in experimental design for computer experiments were explored. The results show that sample size has more impact on metamodel accuracy than the particular experimental design methods utilized, and also that more uniformly distributed sampling does not necessarily lead to more accurate metamodels. Some experimental design methods often appear to be better than the others. |
