Highly efficient photonic convolver via lossless mode-division fan-in

Optical neural networks (ONNs) leverage the parallelism and low-energy consumption of photonic signal processing to overcome the limitations of traditional electronic computing. Optics inherently enables fan-in and fan-out without the Resistor-Capacitor (RC) and Inductor-Capacitor (LC) delays of electrical interconnects. However, for single-mode photonic integrated circuits, reciprocity constraints introduce unavoidable loss during beam combining, hindering large-scale on-chip photonic fan-in. To overcome this challenge, we provide a photonic lossless mode-division fan-in solution for the convolution accelerators. Using inverse design, we developed a compact multimode photonic convolution accelerator (0.42 mm²) with ±15 nm fabrication tolerance and 35 nm optical bandwidth, enabling parallel computation across mode and wavelength dimensions. Experimental results in the C-band confirm a 6-7 bit convolution precision, leading to classification accuracies of 95.2% on MNIST and 87.9% on Fashion-MNIST. Moreover, the device offers a theoretical computational density of 125.14 TOPS/mm², underscoring its potential for scalable and energy-efficient photonic computing accelerators.