Structural optimization concerns the task of determining the best layout of a given device in order to maximize, or minimize, a desired functionality while fulfilling necessary constraints. In entails both size, shape and topology optimization, of which, the latter is the most versatile as it holds the largest degree of design freedom. Topology optimization is already a well-established numerical tool used on every-day-basis in automotive, aerospace, photonics and many other engineering industries. However, the application of these design methods is associated with a significant numerical burden and it is therefore of interest to both academia and industry, that these procedures are accelerated. Accelerating structural optimization solvers can be obtained in many ways, and approaches covers everything from algorithmic development of solvers tailored for massively parallel systems, i.e. utilizing high performance computing; to novel de-homogenization approaches; to the efficient utilization of shared-memory systems such as graphics processing units. In parallel to all these developments is the field of machine learning, which in the past decade has proven a valuable tool in many research disciplines and that the methods are capable of efficiently handing a multitude of different problems. It is therefore obvious that machine learning approaches are being tested for structural optimization problems, and this talk concerns the pros and cons of such applications. Amongst others, the talk will show how neural networks can be used to perform the fine scale de-homogenization task without including any physics, i.e. no stress/strain analysis, in the training process [1]. Another aspect from machine learning that is readily accessible to structural optimization is the use of algorithmic/automatic differentiating and examples from combined shell shape and topology optimization will be shown and the ease of implementation discussed. However, it is important to recognize the limitations of machine learning, as not all aspects are useful for structural optimization problems, and researchers are encouraged to compare new methods to current deterministic state-of-the-art methods [2].