⚙️ Machine-Learning Control for Hardware Systems

Adaptive algorithms for real-time hardware tuning

Overview

This project developed adaptive machine-learning (ML) algorithms for real-time calibration and tuning of analog and mixed-signal hardware systems, including photonic circuits and RF ICs. Traditional manual calibration is time-consuming and cannot adapt to operating environment changes. The ML-based approach achieves automatic, continuous optimization of circuit parameters — pole locations, bias currents, and supply voltages — without human intervention.

Problem

Electronic and photonic circuits exhibit significant parameter drift due to temperature variation, aging, and process spread. Static calibration at production time is insufficient for systems deployed across wide operating conditions. Existing adaptive methods (e.g., gradient descent) are slow to converge and sensitive to local minima. A faster, more robust approach was needed that could run on embedded hardware with limited compute resources.

Approach

A closed-loop AI control system was designed following a Sense → Controller → Actuator → Plant architecture. Sensing hardware monitors circuit performance metrics (e.g., filter center frequency, gain, noise floor). The ML controller processes the sensor data and computes optimal actuator commands in real time. The feedback loop is illustrated below:

Sense ML Controller (AI) Actuator Plant

The ML controller was implemented as a lightweight neural network trained offline and fine-tuned online using Bayesian optimization. A GenAI-assisted engineering workflow accelerated the design of training data pipelines and model architecture search, significantly reducing iteration time.

My Contributions

  • Designed the overall AI control system architecture (sense, decide, actuate) for photonic and RF circuits
  • Developed lightweight ML models (neural networks, Gaussian process surrogates) for online circuit optimization
  • Created GenAI-assisted engineering workflows that accelerated design space exploration
  • Implemented the embedded inference engine on FPGA for real-time deployment
  • Validated the system across multiple circuit types: microring resonators, LC oscillators, filter banks

Results

  • Demonstrated up to 80% faster calibration time compared to conventional gradient-based methods
  • Achieved adaptive real-time tuning across a 100°C operating temperature range
  • Zero manual intervention required after initial system commissioning
  • GenAI-assisted pipeline reduced model design iteration time by approximately 3×

Tools & Stack

  • Python (PyTorch, scikit-learn, GPyOpt) for ML model development
  • FPGA (Xilinx) for real-time inference deployment
  • MATLAB / Simulink for system-level control modeling
  • Custom measurement automation scripts (Python + VISA)
  • GenAI code assistants for design exploration and documentation acceleration

Notes

Some implementation details are withheld. Results described are from prototype characterization and published work.

← Back to Projects