300-Gbps optical interconnection using neural-network based silicon microring modulator

Silicon microring modulators (Si-MRM) are critical components for high-performance electro-optical (E-O) signal conversion at optical interconnections due to their ultrawide bandwidth. However, the current transmission speed at the interconnections is still limited to 240 Gbps because of the low spectral-efficiency, as a result of the inherent modulation nonlinearity of Si-MRMs. Here, we theoretically analyse the modulation nonlinearity of a depletion-mode Si-MRM. Based on the analytical results, we further propose a physics-inspired neural network, named as bidirectional gate recurrent unit (Bi-GRU) to mitigate the signal distortion in Si-MRMs. Bi-GRU matches the analytical E-O modulation dynamics within Si-MRMs, thus can accurately capture the impairment features and accelerate the data transmission speed. We then fabricate a Si-MRM with −3dB E-O bandwidth of 42.5 GHz, achieving an ultrahigh speed optical interconnection with a data rate of 302 Gbps. The maximum spectral-efficiency of modulated signals is improved to 5.20 bit/s/Hz. The results provide insights to develop ultrahigh-speed Si-MRM using emerging AI techniques.