Phase smoothing for diffractive deep neural networks

Optical neural networks have shown promising advantages over its electronic counterpart due to fast computational speed, high throughput, and low energy consumption. For the deep neural networks mimicked by the cascaded phase masks, the conventional training algorithm results in abruptly varying phase profiles. In this work, we show that even two kinds of diffraction formulas give completely different results for the identical rough phase masks, which denotes that the training process harvests the inaccuracy of the used diffraction modeling, which prevents the results from being replicated when the diffraction is formularized by other numerical methods or implemented in the real world. The modeling inaccuracy of each phase mask accumulates to prevent the successful optical implementation. In this work, we propose a two-step training framework to enforce smooth phase masks, and thus reduce the mismatch between numerical modeling and practical deployment.