Achieving non-volatile, low-loss phase modulation with ultra-low energy consumption remains a challenge in photonic in-memory computing. Inspired by electrical memory technologies, mechanisms such as ion-migration, phase change transitions, and ferroelectric polarization have been explored in photonic platforms for memory functions. However, existing materials typically require large device footprints to achieve effective optical index tuning, leading to increased insertion loss and energy consumption. Here, we demonstrate a compact non-volatile phase modulator by incorporating 2D ferroelectric CuInP2S6 (CIPS) into a WS2/CIPS/graphene heterostructure, integrated on a SiN microring resonator. This vertical configuration leverages Cu+-induced polarization in CIPS to electrostatically tune the refractive index of WS2 without introducing additional optical loss or static power consumption. The intralayer Cu+-mediated ferroelectric switching (free from domain wall motion) and high dielectric constant enable the device to operate with an ultra-low switching energy of 2.5 pJ per cycle, a fast write speed of 5 V/µs, and an insertion loss of 0.2 dB. The device further shows stable multi-level (8-bit) memory, with projected retention beyond 10 years. We showcase its potential in photonic in-memory computing by implementing the modulator within an optical neural network, achieving 92% accuracy on the MNIST handwritten digit recognition, establishing new avenues for hardware-accelerated neural networks.
Singh et al. (2026) studied this question.