{ }

FSO Comms with ML-Based Turbulence Compensation

A novel GNN+CNN architecture that reconstructs and classifies optical communication signals from heavily distorted and occluded data, achieving 100% accuracy.

May 2025•completed
FSO Comms with ML-Based Turbulence Compensation screenshot 1
FSO Comms with ML-Based Turbulence Compensation screenshot 2
FSO Comms with ML-Based Turbulence Compensation screenshot 3
FSO Comms with ML-Based Turbulence Compensation screenshot 4
FSO Comms with ML-Based Turbulence Compensation screenshot 5
FSO Comms with ML-Based Turbulence Compensation screenshot 6
View Code
Technologies & Skills
PyTorchMatplotlibDeep LearningCNNGNNComputer VisionQuantization
Project Details

The Challenge: Signal Loss in Optical Communication

This research addresses a critical problem in Free-Space Optical (FSO) communication: atmospheric turbulence distorts signals, making them unreadable. Initial attempts to classify these distorted signals using a standard Convolutional Neural Network (CNN) were unsuccessful, yielding less than 15% accuracy and proving the difficulty of the problem.

An Advanced ML Solution: The GNN+CNN Pipeline

To solve this, the project pivoted to a more robust signal type ("petalled modes") and implemented an innovative two-stage machine learning pipeline:

  1. GNN for Signal Reconstruction: A Graph Neural Network (GNN) was designed to handle the most challenging scenario: severe occlusion, where 75% of the incoming signal data was randomly masked (deleted). The GNN's sole task was to analyze the single remaining quadrant and intelligently reconstruct the full, complete signal pattern.

  2. CNN for Signal Classification: The fully reconstructed signal from the GNN was then passed to a fine-tuned CNN. This network's task was to classify the clean, reconstructed pattern and accurately identify the original data it represented.

Exceptional Results

This GNN+CNN architecture demonstrated remarkable resilience. Even when working with only 25% of the original data, the system achieved a perfect 100% classification accuracy. The success of this model showcases the power of using specialized neural networks not just for classification, but for advanced signal reconstruction in challenging, real-world conditions.

Efficient Inference

The ~300K-parameter model was trained with quantization-aware training, cutting inference compute by about 35% with negligible accuracy loss, and was validated on both CUDA and Apple MPS.