Active learning for efficient nanophotonics inverse design in large and diverse design spaces.

The diverse range of shapes enabled by modern nanofabrication techniques makes it challenging to identify the most optimal design for a desired optical response. While deep learning-based approaches are being increasingly explored for inverse design (especially for optical metasurfaces), they have mostly been limited to small design subsets that constrain the shapes, thicknesses, or parameters like pitch, incident angle, etc. Scalability of such techniques to the full design space accessible to modern nanofabrication remains a challenge due to the difficulty of training models which retain acceptable generalization across wider design spaces. We explore the possibility of using active learning techniques for sample-efficient model training and surrogate-driven inverse design. Specifically, we consider the inverse design of periodic optical metasurfaces with 36 diverse shape classes and broad variations in thickness and pitch. Our results demonstrate that an active learning-driven approach to inverse design can give comparable performance to random-dataset trained models with a substantial (up to 82%) reduction in training dataset generation. The inverse design capability is seen across diverse spectral filter optimization tasks with the added benefit of providing multiple solutions in a single run. As a model- and training-paradigm-agnostic dataset reduction technique, our approach enables deep learning-based inverse design to scale to increasingly larger design spaces.