Gravitational-Wave Coincidence Detection With ML

ISEF Category: Physics and Astronomy

Subcategory: Astronomy and Cosmology  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A gravitational-wave signal can hide below the noise in one detector and still leave a clue in another. That is the whole trick behind coincidence searches. You are not trying to hear one clear shout, you are trying to spot two faint whispers that match. That makes this project a real hunt for buried signals.

What Is It?

Gravitational waves are tiny ripples in space-time from violent events like black hole or neutron star mergers. LIGO and Virgo record huge streams of detector noise, and real signals often look weak or messy. A coincidence search asks a simple question, do two detectors show a matching pattern at the same time?

A graph neural network is a machine learning model that works well when objects connect to each other. In this project, you can treat detector triggers or features as nodes in a graph, then ask the model to decide which pairs or clusters look like real compact-binary events. Think of it like sorting a crowd of overlapping conversations, where you listen for two people repeating the same story at different spots in the room.

Why This Is a Good Topic

This is a strong science fair topic because it has real data, a clear signal-versus-noise problem, and a measurable result. You can test how well your classifier finds weak events, then compare it with a simpler coincidence rule or matched-filter baseline. The project also connects to real astronomy problems, like how researchers find rare events in noisy detector streams. You can learn data handling, feature design, model evaluation, and false-alarm calibration.

Research Questions

• How does a graph neural network compare with a simple coincidence threshold for finding sub-threshold compact-binary candidates?
• What is the effect of changing the graph features on injection recovery in GWOSC O3 data?
• Does adding detector timing consistency improve the classifier's false-alarm rate?
• To what extent do different signal-to-noise ratio bins change recovery efficiency for injected events?
• Which coincidence window gives the best trade-off between sensitivity and false alarms?
• How does the classifier perform on simulated injections from different mass ranges?

Basic Materials

• Computer with enough memory to handle public GWOSC data.
• Python installed with NumPy, pandas, scikit-learn, PyTorch or TensorFlow, and Matplotlib.
• GWOSC public data from the third observing run.
• A notebook environment such as Jupyter.
• Access to published injection sets or the ability to generate simple simulated injections.
• External storage for intermediate data files.
• Basic statistics reference for confusion matrices, ROC curves, and false-alarm rates.

Advanced Materials

• University or cloud computing access with a GPU.
• Python with PyTorch Geometric or a similar graph learning library.
• GWOSC strain data from multiple detectors.
• Published compact-binary injection catalogs.
• Tools for matched-filter or trigger generation.
• Version control such as Git.
• High-capacity storage for feature tables and model checkpoints.

Software & Tools

• Python: Processes GWOSC data, builds features, and trains the classifier.
• Jupyter: Lets you explore detector triggers, plots, and model outputs step by step.
• PyTorch: Trains the graph neural network on coincidence features.
• GWOSC Data Portal: Provides public LIGO and Virgo observing-run data and metadata.
• ImageJ: Not used here, so you should skip it and focus on numerical analysis tools instead.

Experiment Steps

1. Define the exact search problem, such as which detectors, which observing times, and which type of compact-binary events you will target.
2. Choose the trigger or feature format your graph will use, then decide how nodes and edges will represent coincidence.
3. Build a baseline method first, so you have a simple comparison for the machine learning model.
4. Plan an injection-recovery test set that includes both real background and known simulated signals.
5. Set up your evaluation plan, including efficiency, receiver operating characteristic curves, and false-alarm rate calibration.
6. Decide how you will test whether the model keeps working when you change the signal class, detector pair, or noise condition.

Common Pitfalls

• Training on injected signals that are too similar to the test set, which makes the recovery score look better than it really is.
• Mixing data from different observing conditions without tracking detector state, which confuses the classifier.
• Using only accuracy, which hides the fact that rare-event searches care more about missed signals and false alarms.
• Building a graph that ignores timing or detector identity, which removes the very structure the coincidence search needs.
• Skipping background calibration, which leaves you without a real false-alarm rate for your candidate events.

What Makes This Competitive

A strong version of this project does more than train a model. It compares against a clean baseline, uses blind or semi-blind injections, and reports sensitivity at fixed false-alarm rate. You can make it stronger by testing whether the model generalizes across detector pairs, signal masses, or noise periods. Careful calibration and honest uncertainty estimates matter more than a flashy score.

Project Variations

• Try the same coincidence search on binary neutron star injections instead of mixed compact-binary signals.
• Replace the graph neural network with a gradient-boosted tree model and compare sensitivity at the same false-alarm rate.
• Focus on how detector noise state affects recovery by splitting the data into quieter and noisier observing periods.

Learn More

• GWOSC Documentation: Find public LIGO and Virgo data, tutorials, and event resources on the Gravitational Wave Open Science Center site.
• LIGO Open Science Center Papers: Search for overview articles on detector data analysis and transient searches in Classical and Quantum Gravity and related journals.
• MIT OpenCourseWare: Look for courses on signal processing, probability, and machine learning that help with feature design and evaluation.
• NASA ADS: Search for review papers on compact binary coalescences, gravitational-wave detection, and coincidence pipelines.
• PubMed: Not a main source for this topic, but useful if you want to search for statistics or machine learning methods papers linked from broader reviews.
• PyTorch Geometric Documentation: Read the free docs for graph neural network layers, message passing, and example classification projects.