Design Manifolds

Design Manifolds, as first described in external page our 2017 JMD paper, refer to learning a non-linear subspace within a design space that captures meaningful variation among high performing designs. To find these manifolds, we typically learn a non-linear dimension reduction model — such as Kernal PCA, denoised Autoencoders, Variational Autoencoders (VAEs), GANs, or Diffusion Models — that is trained on examples provided by humans, or results from prior optimization solutions. It is the key underlying technique behind some of our follow on work on either Inverse Design or accelerating Optimization methods.