Abstract
Multiscale Topology Optimisation (MTO), by definition, operates over multiple length scales to achieve an optimised structural solution. Consequentially MTO can be computationally very expensive. Machine learning provides the opportunity to accelerate the process utilising surrogate models that approximate the behaviour of the system, allowing for faster evaluation and optimisation. This is particularly advantageous in scenarios where direct simulations at all scales are impractical or prohibitively time-consuming. In this study we demonstrate the effectiveness of a convolutional neural network (CNN) to map topologies of 2D micro-architectures to their corresponding stiffness property matrices. We utilise the topological dataset presented by Jiang et al [1] and generate the stiffness properties using 2D homogenisation [2] for micro-architectures made of a combination of printable polymeric materials (i.e. from a multi-material 3D printer such as an Object 260). We present a successful mapping, achieving results of R 2 greater than 99% and a mean average error of 0.012. Additionally, we examine the transferability of the initial learning of the above trained network. The trained mapping was used on the same topological dataset but with a different material combination. R 2 reduced to 34% and mean average error rose to 0.436. Mappings trained on one material combination will therefore be valid for only a small section of the total design space. To further explore the transferability of CNN, two architectures are proposed and evaluated. One network architecture employs transfer learning and the other incorporates branched connections which account for different printable material combinations. These models are assessed in terms of both the reduction in the computational cost, by leveraging the already trained CNN, and accuracy.