Reducing Computation Complexity in Optimization of Nanophotonic Structures Using Pruning

Inverse design of complex nanophotonic devices is a very computation-consuming task. Deep-learning-based approaches can facilitate this process. However, due to the lack of solid knowledge about the underlying complexity of the input-output relation for a selected class of nanostructures, it is common to select an over-parameterized neural network (NN) for modeling this relation. We present a novel pruning method based on removing weak nodes and connections in the original NN to simplify the input-output relation without imposing significant error. In addition to reducing the model complexity computations, the pruned network can be used to find valuable insight into the physics of device operation. To show the efficacy of our approach, we use it for modeling and inverse design of two classes of nanostructures with different complexities.