Normalization flows for designing metasurfaces

Training data are very expensive for metasurface designs using deep-learning based methods. To achieve data efficiency, we introduced normalization flows to inversely design metasurfaces through learning the probabilistic distribution of the data. To explore the distribution efficiently, we proposed a prior shaping method to re-weight the data based on their fitness to the design target. Through implementing the algorithm, we obtained an emitter with in-band emittance efficiency close to 99%. Our results present a new route for designing metasurfaces inversely.

Duke Scholars

Author Willie John Padilla Electrical and Computer Engineering