Accelerated Optimization of Robust Nanophotonic Devices via Deep Learning

The potential for nanophotonic devices to disrupt existing and create new commercial applications has led to a surge in design and manufacturing research in recent years. Yet, new technologies must not only demonstrate performance advantages over legacy solutions, but also improved fabrication cost and reliability. Therefore, inverse-design techniques which optimize guaranteed performance in the presence of fabrication uncertainties are needed to maximize the yield achievable within a given process window. However, simulating a nominal structure and all potential perturbations caused by a variety of highly complex and coupled fabrication uncertainties results in a combinatorial explosion of solutions that ultimately makes direct optimization intractable. To overcome this, we exploit deep learning and demonstrate a neural network that accurately predicts the performance of a representative metasurface supercell in the presence of fabrication uncertainties. The trained neural network is subsequently paired with a modified multi-objective optimization procedure which enables one to study the tradeoffs between nominal performance and guaranteed performance.