Computationally efficient and fabrication error tolerant inverse-designed mode converters and mode-division (de)multiplexers

Le résumé fourni par la source

The explosion in data processing, storage, and communication of present time demands for consistent growth of cloud services and supporting hardware technologies. Silicon photonics (SiPh) offers higher data transmission and processing speed at low latency and low thermal dissipation. One of the current research drives in SiPh is to achieve dense integration of photonic interconnects. Computational inverse design techniques have shown potential to become reliable means for designing compact nanophotonic devices. Much effort has been made to obtain final designs that are robust to fabrication imperfections. In this work, we experimentally demonstrate optical mode converters (MCs) on the silicon-on-insulator platform designed using the computationally efficient shape optimization method. The mode conversion is performed between different transverse electric (TE) mode pairs among TE0, TE1, TE2, and TE3 modes. These MCs have mode conversion efficiencies above 95%, and the insertion loss ranging from 0.3 dB to 1 dB over a wavelength span of 80 nm ranging from 1.5 μm to 1.58 μm. Maximum modal crosstalk found experimentally in the C-band is -19 dB. The conversion efficiency drops at most by 2.2% at 1.55 μm for 10 nm over/under etch, implying good robustness to dimensional variations. To characterize their performance in the time domain, a 28 Gbps on-off keying (OOK) and a 20 GBaud pulse amplitude modulation (PAM-4) payload transmissions were performed, which supports their utility for high throughput data communications. The open eye diagrams exhibit Q-factors of 8 dB. Additionally, the mode conversion mechanisms of these MCs are investigated by studying the simulated electromagnetic field patterns and validated by supportive data. This interpretation of the working principles leads to formulating the optimization such that more efficient designs are obtained without requiring larger design area. It also guides to a two-step design approach for dealing with relatively more complex design problems.Inverse design usually deals with many design parameters, and the optimized designs are eventually the local optima in the non-convex design parameter space. But a local optimum may not turn out to be a good design with acceptable performance. A collection of good designs is very useful, since one design may turn out to be superior to the rest for an application-specific performance attribute. An exhaustive search for all the good designs in a given parameter space is computationally burdensome. To reduce the computation cost involved with the 3D optimization of any type of optical interconnect which can be designed in the shape optimization method, a machine learning-based regression model has been proposed and implemented to the TE0-TE1 mode converter design for a demonstration. It shows a reduction of the computation load by 35% in the 3D optimization step, which is the most computationally expensive part.Finally, using the density topology optimization technique, several three-channel mode-division (de)multiplexers (MDMs) are designed with a design footprint of 4.5×4.5 μm2. Experimental results show maximum insertion loss of 1.2 dB and channel crosstalk below -18 dB in the C-band (1.53 – 1.565 μm). A physics-guided approach is adopted for faster convergence to the optimum designs. An aware selection of the initial design parameters shows some advantage over the random selection of the starting conditions. However, the optimized designs are sensitive to fabrication imperfections. The channel crosstalk increases by several folds after fabrication (~6 dB on average)

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.