Nonvolatile Programmable Fiber Weight Bank for Convolution Calculation

Fiber optics are a reliable medium for sensing and transmission, but processing these optical data requires switching to the electrical domain, which introduces large latency and thermal footprints. Photonic convolutional neural network (PCNN) offers the capability to directly process optical data and exhibits the characteristics of high parallelism, ultrafast processing speeds, and low energy consumption. However, existing PCNN schemes are limited by additional controllers, high-precision configuration, and complex alignment devices. A potential solution is to design a programmable fiber weight bank for PCNN, which supports all-optical in-fiber calculation. Here, we integrate Ge1.4Sb1.22Se3.96Te1 (GSST) on the lateral facets of commercial seven-core optic fibers as a programmable fiber weight bank. The coupling efficiency between the GSST and fiber is enhanced by employing side-polishing techniques. Each GSST cell can be optically altered to exhibit a 16-level weight, depending on the power of the light pulse within the fiber. The fiber weight bank has reconfigurability, parallel computing capability, and can be easily scaled. In addition, a fiber-PCNN is constructed to demonstrate digital recognition with an accuracy of 90.0% (theoretically 91.6%), which combines the calculative capabilities of PCNN with the merits of a fiber, thereby providing a novel solution for remote all-optical computing.