Learning Guide

Warning

This guide is under construction. The content below outlines what will be available here once the guide is complete.

The Learning Guide is designed for three audiences: complete beginners with no prior background; physicists who want to understand the ML side of Sage; and machine-learning practitioners who want to understand the GW physics. See Guide for Complete Beginners if you are starting from scratch.

What Will Be Here

GW Physics Primer (for ML readers)

A self-contained introduction to the gravitational-wave detection problem — no prior GW knowledge assumed. Topics will include:

  • What compact binary coalescence signals look like in time and frequency

  • Why detector noise is coloured and what whitening achieves

  • How matched filtering works and why it is the optimal linear detector

  • What false alarm rate means and why it is the correct figure of merit

  • A plain-language explanation of the MLGWSC-1 benchmark setup

ML Primer (for GW readers)

A concise introduction to the supervised-learning concepts used in Sage — aimed at physicists already comfortable with GW data analysis. Topics will include:

  • What a convolutional neural network learns from time-series data

  • Why on-the-fly data generation matters (and what data-reuse bias looks like in practice)

  • How binary cross-entropy connects to detection statistics

  • What torch.compile does and why it is safe to treat as a black box

Step-by-Step Tutorials

Notebook-style tutorials that build from first principles to a full training run:

  1. Downloading and inspecting O3b noise data

  2. Generating your first IMRPhenomPv2 waveform

  3. Running the whitening and multirate pipeline on a single sample

  4. Training a minimal ResNet on synthetic CBC data

  5. Evaluating a checkpoint against the MLGWSC-1 injection set

All tutorials will be available as Google Colab notebooks for zero-installation use.

Worked Examples

Case studies showing how to adapt Sage for specific research questions, including:

  • Changing the mass prior to target a different population

  • Swapping the frontend for a different time-frequency representation

  • Evaluating on a custom noise dataset

Exercises

Short, self-contained exercises to test understanding — with solutions provided.