Fast, Efficient Metamaterial Device Design via Neural Networks

Stanford researchers in the Fan Lab have developed a method that dramatically accelerates and optimizes metamaterial design with little computational resource and time using generative neural networks. The network uses known high efficiency device designs as training data to learn features, and produces device designs that mimic high performance devices. The generated devices are quickly characterized with an electromagnetics solver, and the fraction of the devices that are found to be high efficiency are kept. The network further optimizes the topology to produce even higher efficiency, robust designs. These new devices can be fed back into the neural network as additional data for training. The method is particularly suited to optimize complex, aperiodic topology designs for high performance optic devices, such as a flat lenses, gratings, and filters, and potentially electromagnetic devices, such as antennas and impedance surfaces.

Stage of Development – Prototype
The Fan Lab uses the methodology for ongoing device design and research.