Inverse Design of Unidirectional Transmission Nanostructures Based on Unsupervised Machine Learning

Structural design is an important driving force for technological development in nanophotonics. To achieve better performance of nanophotonic devices, the freedom of structural design space needs to be expanded. Unsupervised learning algorithm in deep learning provides a great platform to expand design freedom of structures and avoid the "curse of dimensionality" effectively. It performs well on extracting important features from high-dimensional data and excavating potential rules. In this work, an unsupervised convolutional neural network is built to inversely design nanostructures with unidirectional transmission. Near-field information with high dimensions is recognized and extracted into a 2D feature space which maintains high physical continuity and maps to far-field transmittance effectively. The feature space is further expanded to the whole space by optimistic Bayesian multisampling, from which nanostructures with transmittance over 95% forward while less than 40% backward are inversely designed. Moreover, the relation between near-field information and far-field transmittance is explored. A feasible design method of nanostructures is proposed based on unsupervised learning with design space expanded. This design mentality exhibits a way of extracting near-field features to analyze far-field spectra with deep learning algorithms, which is suitable for more abundant physical design and can be extended to other similar systems.