抄録
Focusing on nanophotonic power splitters, we show that a new class of generative neural network, so
called an adversarial conditional variational autoencoder with cycle consistency, can design a series of
devices that achieve nearly arbitrary target performance, with an excellent capability to generalize training
data produced by the adjoint method. This method is generic and is expected to be applicable to a
broad range of design problems, not limited to nanophotonics.
