Digital-twin based damage-resilient programming of classical and quantum operations on a photonic processor

Sammanfattning

This thesis investigates whether a programmable photonic integrated circuit can remain useful even after some of its thermo-optic phase shifters have become permanently disabled. The approach is based on a digital twin: a differentiable model of the chip is trained from experimental current-output data and then used to solve an inverse-programming problem to recreate desired optical intensity transformations. The method is demonstrated on an 8 × 8 photonic circuit for both classical and quantum tasks. The trained twin accurately modeled the experimental setup, achieving a test fidelity greater than 0.99 for all disabled-phase-shifter configurations. Using the twin as a differentiable forward model to solve for currents, a mean fidelity of 0.9914 ± 0.0001 was measured over 100 target intensity transformations of the form T=|U|², where U is a Haar-random unitary. When any of five individual phase shifters were disabled this mean fidelity remained above 0.9739 ± 0.0002, and under cumulative disabled phase shifters the fidelity decreased to 0.8100 ± 0.0006 at 28 disabled phase shifters. A postselected dual-rail CNOT gate was also implemented and the highest postselected fidelity of 0.8730 ± 0.0097 was obtained for 20 disabled phase shifters, indicating that damage could act as a regularizer for two-photon interference.