Low-Loss and Nonvolatile Multilevel Phase Shifter for Optical in-Memory Computing

We present a low-loss and nonvolatile multilevel phase shifter, designed to achieve precise and energy-efficient phase modulation. By incorporating both coarse and fine tuning blocks, our phase shifter offers 121 distinct phase-shifting states, with an exceptionally low insertion loss of $0.039 \text{dB}/\pi$ as demonstrated in simulations. To validate the performance of our proposed phase shifter, we utilize it in Mach-Zehnder Interferometer networks and base on this to construct optical convolution kernels for a Convolutional Neural Network, which is tested using the MNIST handwriting digit database and achieves an outstanding accuracy rate of over 97% in handwriting digit recognition tasks.