Deep Inverse-Designed Nanophotonic Processors with Structural Nonlinearity from Repeated Phase Encoding

Abstract

Passive nanophotonic scattering regions implement linear optical transformations, and cascading input-independent regions alone does not create functional depth because the resulting transformations collapse into a single linear operator. Here, we introduce repeated phase encoding between inverse-designed passive transformations to generate an input-conditioned multilayer optical map without interlayer photodetection. Each re-encoding introduces additional phase-dependent optical pathways, producing structural nonlinearity with respect to the encoded variables while every scattering region remains passive and linear in the optical field. Under a controlled MNIST depth sweep, classification accuracy increases from 83.53\% with one layer to 93.61\% with seven layers, whereas the input-independent passive control saturates at 86.34\%. The depth trend also persists in a time-multiplexed CIFAR-10 patch model. We further realize three jointly trained $16\times16$ transformations, each independently implemented as an inverse-designed nanophotonic region, with relative complex transmission errors of 7.63\%, 7.59\%, and 8.80\%. The reconstructed electromagnetic stack reaches 91.79\% accuracy after phase calibration, compared with 91.95\% for its surrogate model. These results establish repeated input encoding as a route to functional depth in compact inverse-designed nanophotonic processors.