Development\nof Photonic In-Sensor Computing Based\non a Mid-Infrared Silicon Waveguide Platform

Neuromorphic\nin-sensor computing has provided an energy-efficient\nsolution to smart sensor design and on-chip data processing. In recent\nyears, various free-space-configured optoelectronic chips have been\ndemonstrated for on-chip neuromorphic vision processing. However,\non-chip waveguide-based in-sensor computing with different data modalities\nis still lacking. Here, by integrating a responsivity-tunable graphene\nphotodetector onto the silicon waveguide, an on-chip waveguide-based\nin-sensor processing unit is realized in the mid-infrared wavelength\nrange. The weighting operation is achieved by dynamically tuning the\nbias of the photodetector, which could reach 4 bit weighting precision.\nThree different neural network tasks are performed to demonstrate\nthe capabilities of our device. First, image preprocessing is performed\nfor handwritten digits and fashion product classification as a general\ntask. Next, resistive-type glove sensor signals are reversed and applied\nto the photodetector as an input for gesture recognition. Finally,\nspectroscopic data processing for binary gas mixture classification\nis demonstrated by utilizing the broadband performance of the device\nfrom 3.65 to 3.8 μm. By extending the wavelength from near-infrared\nto mid-infrared, our work shows the capability of a waveguide-integrated\ntunable graphene photodetector as a viable weighting solution for\nphotonic in-sensor computing. Furthermore, such a solution could be\nused for large-scale neuromorphic in-sensor computing in photonic\nintegrated circuits at the edge.