🔬 Automatic Tuning of Silicon Photonic Filters

Machine-learning-driven control systems for photonic resonators

Overview

Silicon photonic filters, such as microring resonators and Mach-Zehnder interferometers, are highly sensitive to temperature and fabrication variation, requiring continuous automatic tuning to maintain performance. This project developed machine-learning-driven control systems that stabilize photonic filter responses in real time, enabling robust signal filtering and adaptive interference rejection for next-generation photonic transceivers.

Problem

Photonic resonators drift with thermal variation at rates that make manual or open-loop tuning impractical. Traditional PID controllers struggle with the nonlinear, hysteretic response of thermo-optic actuators. The challenge was to achieve sub-microsecond locking across wide temperature ranges while rejecting strong out-of-band jamming signals without human intervention.

Approach

A closed-loop architecture was designed using on-chip power monitors as feedback sensors, a compact ML inference engine as the controller, and thermo-optic phase shifters as actuators. An inline SVG diagram of the control loop is shown below:

RF In Modulator Photonic Filter Photodetector RF Out

The ML controller was trained offline using characterization data and deployed as a lightweight inference engine co-packaged with the photonic chip. Gradient-free optimization and Bayesian methods were compared; a custom online adaptation algorithm was selected for production.

My Contributions

  • Designed and validated the closed-loop control architecture for silicon microring resonators
  • Developed and trained the ML inference models for online filter stabilization
  • Implemented the thermo-optic actuator driver circuitry and feedback sensor interface
  • Co-authored peer-reviewed publications reporting experimental results
  • Characterized prototypes across temperature ranges from −40°C to +85°C

Results

  • Demonstrated up to 10× faster locking speed compared to conventional gradient-descent methods
  • Achieved >40 dB out-of-band jammer rejection in characterization
  • Real-time stabilization maintained across a 100°C operating range
  • Algorithm reduced required actuator power by approximately 30%

Tools & Stack

  • Silicon photonic test platform (microring resonators, MZI filters)
  • Python (NumPy, SciPy, scikit-learn) for model training and simulation
  • MATLAB / Simulink for control system modeling
  • Custom FPGA firmware for real-time ML inference
  • Keysight VNA and optical spectrum analyzer for characterization

Notes

Some implementation details and chip-level schematics are withheld. Published results are available on Google Scholar.

← Back to Projects