Integrating material selection with design optimization via neural networks

The engineering design process often entails optimizing the underlying geometry while simultaneously selecting a suitable material. For certain class of simple problems, two are separable where, for example, one can first select an optimal material, and then optimize geometry. However, in general, not separable. Furthermore, discrete nature material selection is compatible with gradient-based optimization, making simultaneous optimization challenging. In this paper, we propose use variational autoencoders (VAE) optimization. First, data-driven VAE used to project database onto continuous differentiable latent space. This coupled fully-connected neural network, embedded finite-element solver, neural-network’s built-in gradient optimizer back-propagation exploited during proposed framework demonstrated using trusses, where needs be chosen from database, cross-sectional areas truss members. Several numerical examples illustrate efficacy framework. Python code these experiments available at github.com/UW-ERSL/MaTruss.