Machine learning based Topology Optimization

Sammanfattning: This thesis aims to find a design suggestion for the chassis of a new micromobility vehicle developed by the company LEVTEK SWEDEN AB by using a multiple load, compliance based topology optimization. For this purpose, 9 static load cases are suggested, 4 of which have been derived from dynamic scenarios using an equivalent static load inspired approach based on free-body diagrams. The results from the topology optimization is presented as a design suggestion, but further post-processing is needed. Additionally, the extent of which machine learning could be applied for speeding up of the topology optimization was explored, and it was concluded to be feasible on a 2D cross section of the deck given the state-of-the-art and available resources. For this purpose a convolutional neural network proposed by Sosnovik & Oseledets (2017) was used, which demonstrated strong potential for learning a specific design domain, and it was investigated if the close-to-optimal solutions found by the network could be used as an initial guess for further topology optimization. It is concluded that transferability and consistency needs to be further investigated for this deep-learning approach.