Ultra-compact Photonic Encoder for Optoelectronic Fusion Computing

Abstract Photonic encoders (PEs) provide high-speed optical front ends for optoelectronic fusion computing by transforming input data into compact feature representations before electronic processing. However, the large footprint of existing photonic encoder architectures limits the integration density required for high-dimensional data processing. Here, we propose an ultracompact photonic encoder (UC-PE) implemented on a standard silicon-on-insulator platform, leveraging a fabrication-aware inverse-design strategy. The footprint of the UC-PE is only 33.28 μm2, and the device works as a fixed passive optical encoder without task-specific photonic reconfiguration. The UC-PE is integrated into a hybrid optoelectronic computing system, where the optical front end generates compact feature maps and the electronic backend performs task-dependent learning. For image classification, the system achieves 94.29% accuracy on the MNIST data set and maintains classification capability on texture-rich Fashion-MNIST data set without photonic hardware reconfiguration. Beyond classification, the system performs image compression and reconstruction with a mean PSNR of 28.53 dB. These results indicate that the UC-PE can reduce the footprint barrier of photonic encoders and provide a compact route toward high-throughput optoelectronic fusion computing.