Structural optimisation in this case refers to the branch of engineering in which a physical structure is optimised with respect to certain performance criteria. Finite element methods are one technique to facilitate structural optimisation whilst also allowing visualisations and acting, generally, as a good design tool.
In this paper by Hanna and Mahdavi, a machine learning approach is applied to a model under a physical simulation. Based on the premise that natural optimisation (in evolution, for example) occurs by trial and error, and is therefore painstakingly slow, a method which employs some ability to learn from past behaviour is proposed.
Past experience in engineering has shown that practiced engineers are good at finding near-optimal solutions based on learning from experience (and a knowledge of physics), and this paper shows that, for a well-defined structural design problem, a near optimal solution can be found using the machine learning approach proposed.
In this case, traditional optimisation techniques are used to provide the training data. The main contribution of this work is its success in finding a good solution when other optimisation procedures fail. Alternative methods may fail, for example, when there is a local but not global optimal solution is found, whilst the proposed method benefits from enough experiential learning to overcome this.
Finally, the results presented in the paper show an improvement in computation time over other optimisation procedures suggesting that they could well be a very practical addition to the design pipeline.
In terms of its application, this paper refers to structural optimisation of lattice-like materials under stress but in principal, the idea of structural optimisation expands much further and to a much larger scale of structure. The idea of machine learning for optimal structures in the built environment is one under much current investigation and still remains a significant challenge.