Nonvolatile silicon photonics based on monolithic back-end-of-line integration phase-change-materials platform for optical neural networks

Benefiting from high parallelism and low latency, photonic integrated circuits (PICs) constructed from on-chip building blocks with diverse functions have emerged as a promising technology in the realm of optical neural networks (ONN). Tunable components, through the utilization of physical mechanisms such as thermo-optic effect and free-carrier plasma dispersion effect, structural motion like microelectromechanical systems (MEMS), or material properties including liquid crystal and two-dimensional materials, play a pivotal role in enabling reconfigurability within PICs. Among these reconfiguration schemes, chalcogenide phase change materials (PCMs) based photonic devices have attracted extensive attention owing to their high energy efficiency and integration density brought by huge refractive index contrasts and nonvolatility of PCMs. However, this nonvolatile modulation method meets difficulty in scalability since the process flow of integrating PCMs into silicon photonics is insupportable in the foundries. Here, we demonstrated a back-end-of-line (BEOL) integration platform for the monolithic integration of PCMs into silicon photonic devices without modification in standard process design kits (PDK). This is achieved by fabricating a low-loss oxide trench to expose the waveguide core at the functional area from the top dielectric layer, with assistance from a silicon nitride etch stop layer. On this basis, integrated photonic devices with stable switching performance and repeatable multi-bit storage capability have been developed, possessing the potential for crucial blocks of PICs in ONN applications that require infrequent reconfiguration, such as hardware error correction before training and data storage in pre-trained models.