Local Variance-Based Calibration of Programmable Photonic Interferometer Meshes

Technical Summary: Local Variance-Based Calibration of Programmable Photonic Interferometer Meshes

Problem Statement
Programmable photonic interferometer meshes, composed of Mach–Zehnder interferometers (MZIs) and phase shifters, serve as universal platforms for reconfigurable linear optical transformations. However, their performance is critically dependent on the accurate calibration of these components. Conventional calibration methods typically rely on strict node isolation (keeping one input dark), dedicated routing paths, orthogonal training states, reference channels, or prior phase–voltage characterization. These requirements become increasingly difficult to satisfy as the scale of thermally tuned meshes grows. Fabrication imperfections, thermal drift, and crosstalk introduce phase and amplitude errors that degrade stability. Existing alternatives, such as self-configuring frameworks based on orthogonal inputs or temporal averaging, often demand complex hardware, mutual coherence, or extensive upstream phase sweeps, limiting their practicality for large-scale, dynamically reconfigurable systems.

Methodology
The authors introduce a local variance-based self-calibration method that relies solely on intensity measurements and does not require node isolation, reference paths, or prior knowledge of the phase–voltage response of thermo-optic actuators. The core principle involves applying controlled phase perturbations to a tunable element and analyzing the statistical variance of the measured output intensity.

The method operates on two distinct calibration signatures derived from the interferometric transfer function:

1. MZI Calibration: When coherent fields are injected into the two inputs of an MZI, the output intensity variance follows a characteristic $|\sin \theta|$ dependence, where $\theta$ is the internal phase. The minima of this variance (at $\theta = 0$ and $\theta = \pi$) correspond to the "bar" and "cross" operating states. Crucially, this relationship holds even when using a finite, non-calibrated ensemble of random phase settings (e.g., $M=5$ random voltage values) rather than a continuous $0$–$2\pi$ sweep. The random sampling rescales the curve but preserves the location of the minima.
2. Phase Shifter (PS) Calibration: For standalone phase shifters operating under balanced interference (where the preceding MZI is set to a 50/50 splitting ratio), the output variance with respect to the internal MZI phase follows a $|\cos \phi|$ dependence, where $\phi$ is the phase shifter's setting. The minima occur at the quadrature points ($\phi = \pi/2, 3\pi/2$).

The experimental procedure is a fully automated, two-stage process applied to an $8 \times 8$ silicon-nitride processor:

• Stage 1: MZIs are calibrated sequentially from the output layer toward the input. Once an MZI is identified and fixed at a bar or cross state, it provides a deterministic routing path for calibrating the preceding layer.
• Stage 2: Phase shifters are calibrated under balanced-interference conditions, utilizing the same variance-minimization principle with a fixed set of random auxiliary phase settings.

Key Results
The method was validated on an $8 \times 8$ Clements mesh containing 28 MZIs and 28 phase shifters.

• Calibration Accuracy: The variance-based approach successfully identified bar/cross states for MZIs and quadrature points for phase shifters. Residual phase errors were found to be within approximately $\pm 0.2$ rad for MZIs and $\pm 0.1$ rad for phase shifters across the full tuning range.
• Robustness: The method functioned effectively starting from arbitrary, unknown initial phase settings and using only a small, non-calibrated set of random voltage values ($M=5$) per element.
• System-Level Performance: To demonstrate practical utility, the authors implemented an embedded $4 \times 4$ Hadamard transformation on the fully calibrated $8 \times 8$ processor. When compared to the manufacturer-provided factory calibration, the variance-based calibration resulted in significantly improved output uniformity and reduced temporal fluctuations across all channels.

Significance and Claims
The paper establishes local output variance as a simple, intensity-only calibration observable for programmable photonic meshes. The authors claim that this approach offers a practical alternative to existing strategies by eliminating the need for:

• Conventional node isolation or diagonal routing.
• Mutually orthogonal or coherent multi-port inputs.
• Prior phase–voltage characterization or calibrated actuator sweeps.
• Dedicated reference channels or embedded monitoring hardware.

The method is described as a "calibration primitive" that is compatible with discrete random phase ensembles. It retains the compact Clements topology (unlike architectures requiring auxiliary ports for independent access) while replacing the requirement for isolated transmission minima with a statistical signature. The authors conclude that this approach provides a scalable, low-overhead solution for stabilizing large-scale, thermally tuned, and dynamically reconfigurable photonic processors, making it suitable for applications in optical computing, neuromorphic processing, and quantum information processing where robust, automated calibration is essential.