Multi-Task Learning for Forward Prediction of Dispersion Diagrams in 2D Acoustic Metamaterials

Metamaterials are unique solid structures which can cease the transmission of waves within a specified frequency range. They have widespread application in critical fields like ultrasound imaging, wearable health devices and others. The topological design of metamaterials is a computationally demanding process and often requires human supervision of domain experts. In this paper we present a deep learning based computer vision architecture to predict the dispersion relation vectors from metamaterial topology images, referred as forward prediction problem. The novel contribution of our work is a methodology to use multi-task learning along with transfer learning for forward modeling. We propose to use pretrained vision models with multiple heads, one for each of the dispersion branches. We demonstrate through experimental results that it is possible to train such a network and do so efficiently.