Emerging methods for Topology Optimisation

This thesis presents the culmination of research into the question of how emerging technologies in computer science can be utilised to improve topology optimisation in terms of efficiency and solution quality. The contained work has especially focused on alternatives that reduce the need for high-performance computing resources without loss in structural detail. Overall, the goal and motivation behind this project was to both assess the current state of this increased multidisciplinarity and investigate future potential avenues.

Similarly to most other computational disciplines, the adoption of artificial intelligence in topology optimisation related research has increased immensely in recent years. An extensive portfolio of published articles proposing different applications of neural networks to improve the computational efficiency of topology optimisation now exists. This segment of the literature does, however, demonstrate some of the main caveats in benchmarking and transparent assessment of such frameworks. Consequently, there is a trend of continued work on artificial intelligence applications in topology optimisation that are fallacious and intractable. As a part of this thesis, a review of the current state-of-theart artificial intelligence applications within the field of topology optimisation investigates and exemplifies the ramifications of these factors. In addition to
providing an elaborate assessment of the existing literature, the presented review also describes important factors in designing and benchmarking future applications.

During the review, it was found that a neural network trained for performing dehomogenisation was one of the more promising applications. This approach offers a more efficient approach to dehomogenisation compared to conventional methods, that is also more stable in the presence of orientation singularities. This gain in efficiency does, however, prove to come at a trade-off in terms of solution quality. Dehomogenisation has an inherent parallel to texture mapping from computer graphics, and the remaining part of this thesis investigates how the specific technique of phasor noise for pattern generation can be utilised as a more deterministic and controllable approach to efficient dehomogenisation. Results promote this new heuristic approach as an improvement in terms of the efficiency-quality trade-off, compared to existing methods. There are still some inherent challenges associated with the phasor-based approach, such that further developments of this concept or integration of emerging technologies from other scientific fields, are highly relevant avenues for continued work in multi-scale topology optimisation.