Calibratable 90° hybrid based on a reconfigurable on-chip optical neural network

Area of Science:

• Photonics and Optical Engineering
• Integrated Optics
• Artificial Intelligence Hardware

Background:

• Silicon-on-insulator (SOI) platforms are crucial for integrated photonic devices.
• Reconfigurable optical neural networks (ONNs) offer programmable functionalities.
• Fabrication variations and packaging complexities challenge the performance of optical systems.

Purpose of the Study:

• To implement a 90° optical hybrid using a reconfigurable optical neural network (ONN) on an SOI platform.
• To enhance robustness against fabrication variations and packaging imbalances.
• To demonstrate adaptive compensation capabilities of the ONN architecture.

Main Methods:

• Cascaded four-port multimode interferometers (MMIs) were used to build the programmable ONN.
• The network was trained to match target system transfer matrices.
• Simulations and experiments were conducted to evaluate performance under various conditions, including introduced error matrices and output coupling imbalances.

Main Results:

• The ONN achieved an average phase error of 0.26° and CMRR of -30.8 dB, even with introduced MMI errors.
• The network adaptively compensated for output port coupling imbalances, reducing packaging complexity.
• Simulations showed average worst-case phase error of 0.24° and CMRR of -34.2 dB, with no significant degradation due to coupling imbalance.
• Experimental results showed a maximum phase error below 1.32° and CMRR better than -36.6 dB after training.

Conclusions:

• The developed ONN on SOI provides a robust and adaptable solution for 90° optical hybrids.
• The programmable nature of the ONN effectively mitigates device-level imperfections and packaging challenges.
• This approach significantly simplifies optical packaging while maintaining high performance.