In the current state of the art, a direct shape synthesis is not possible, the main obstacle being the combinatorial nature of an associated optimization problem which is non-polynomial in time. On the other hand, there is a strong evidence that a skilled designer can provide designs with performance close to fundamental bounds...
How close can we go with clever, albeit brute-force techniques? Can we go closer? Is it possible to improve the design performance with the knowledge of the first differences or utilization of machine learning? To address these questions, a novel memetic framework combining local and global optimization routines is developed, combining the advantages of an adjoint formulation of topology optimization and of an evolutional algorithm. Various geometry- and topology-based metrics like shape regularity are being incorporated as well.
Figure: One run of topology optimization based on the exact re-analysis.
