⚛️ general relativity

Normalizing Flow Emulator for Gravitational Wave Detection Probability in the LIGO/Virgo/KAGRA O4 Observing Run

This paper presents a machine learning-based Normalizing Flow emulator that efficiently and accurately estimates gravitational wave detection probabilities for the LIGO/Virgo/KAGRA O4 observing run, offering a faster and more precise alternative to traditional Monte Carlo injection campaigns for population analysis.

Original authors: Julius Gassert, Ignacio Magaña Hernandez, Ariel J. Amsellem, Antonella Palmese, Barnabas Póczos

Published 2026-10-02

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Julius Gassert, Ignacio Magaña Hernandez, Ariel J. Amsellem, Antonella Palmese, Barnabas Póczos

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The universe speaks in riddles, but sometimes it whispers in ripples. When massive objects like black holes or neutron stars collide, they send out gravitational waves—tiny distortions in the fabric of space and time that travel across the cosmos to reach us. To hear these whispers, scientists use massive, ultra-sensitive instruments called interferometers, such as the ones operated by LIGO, Virgo, and KAGRA. These detectors are incredibly powerful, yet they have a limitation: they can only "hear" the loudest events. A collision that happens far away or involves smaller objects might be too faint to register, while a nearby, massive crash is clear as a bell. This creates a bias in what we see; our catalog of cosmic events is naturally skewed toward the loud and the close. To understand the true nature of the universe, scientists must correct for this bias, figuring out exactly how likely their detectors are to catch any given event. This process, known as determining the selection function, is essential for turning a list of detected collisions into a reliable map of the cosmos.

For decades, the standard way to solve this problem has been a method called Monte Carlo injection. Imagine trying to understand how a fishing net catches fish by throwing thousands of fake fish into the water and seeing which ones get caught. In the world of gravitational waves, researchers simulate millions of potential collisions, add realistic noise to mimic the detector's environment, and run them through the same search software used for real data. By counting how many of these simulated events are "found," they can estimate the probability of detection for any set of conditions. However, this approach is incredibly expensive. It requires vast amounts of computing power and time, and it struggles when scientists want to test very specific or unusual types of cosmic populations. If a population has a sharp, narrow feature—like a specific mass that is very common but rare elsewhere—the standard method might miss it entirely because it doesn't have enough simulated samples in that exact spot.

A team of researchers has now introduced a faster, more flexible alternative using a type of machine learning called a Normalizing Flow. Instead of just counting hits and misses like a simple tally sheet, this new approach learns the shape of the data itself. The researchers trained a computer model on the results of a massive simulation campaign designed for the upcoming fourth observing run of the LIGO/Virgo/KAGRA network. This campaign involved simulating roughly 433 million potential events, of which about 1.5 million were kept for detailed study, with around one-third of those being successfully detected. The machine learning model studied the parameters of the detected events—such as the masses of the colliding objects, their distance, and their spin—and learned the complex, multi-dimensional patterns that make an event detectable. Once trained, this model acts as an emulator, a digital twin that can instantly calculate the probability of detecting any new event without needing to run a fresh, expensive simulation.

The results of this work show that the emulator is remarkably accurate. When the researchers tested the model by generating new, fake events based on the same rules used in the original simulations, the distribution of those fake events matched the real data almost perfectly. The model successfully reproduced the complex relationships between different properties, such as how the mass of the objects relates to their distance and spin. In tests comparing the emulator's predictions against the traditional, slow method of re-sampling, the two agreed closely. However, the emulator showed a distinct advantage when dealing with population models that have sharp or peaked features. In these cases, the traditional method often produced unreliable results because it lacked enough samples to probe the narrow features accurately, leading to large uncertainties. The emulator, by contrast, provided precise estimates even for these difficult cases, effectively smoothing out the statistical noise that plagues the older technique.

There is a minor caveat to this success. In a very small fraction of cases—less than one-quarter of one percent—the model predicted a detection probability slightly higher than 100 percent. This is physically impossible, as nothing can be more certain than certainty. However, the researchers found that these over-predictions only occurred in regions of the parameter space where detection is already virtually guaranteed, such as for very massive objects nearby. The error was small and confined to areas where it mattered least, so the team simply clips any value above 100 percent down to 100 percent. This tiny imperfection does not undermine the utility of the tool. The researchers have made their code and the fully trained models publicly available, allowing other scientists to quickly evaluate detection probabilities for their own studies.

This new tool represents a significant shift in how gravitational wave science handles data. By moving from a brute-force counting method to a learned density model, the team has created a system that is not only faster—capable of evaluating a million events in seconds on a standard laptop—but also more robust against statistical pitfalls. It allows scientists to explore a wider variety of cosmic scenarios with greater confidence, ensuring that the map of the universe they build is not just a picture of what their detectors happened to catch, but a true reflection of what is out there. As the next observing run begins, this emulator will serve as a crucial guide, helping to correct for the biases of the instruments and revealing the true population of colliding black holes and neutron stars across the cosmos.

Technical Summary: Normalizing Flow Emulator for Gravitational Wave Detection Probability

Problem Statement
Accurately estimating the detection probability of gravitational wave (GW) events, given their source parameters, is a critical component of GW data analysis and population inference. Traditionally, this requires expensive Monte Carlo injection campaigns where simulated signals are injected into detector noise and processed through search pipelines. These campaigns generate a discrete set of detections and non-detections in a high-dimensional parameter space. While useful, this approach is computationally intensive and laborious. Furthermore, when estimating detection efficiency for population models with sharp or peaked features (e.g., specific mass or spin distributions), traditional Monte Carlo importance sampling (MC-IS) suffers from a reduced effective number of samples ( $N_{eff}$ ). This reduction leads to high variance in the integral approximation, potentially biasing population inference results.

Methodology
The authors propose a machine learning-based approach using Normalizing Flows (NFs) to emulate the detection probability $P(D|\theta)$ for a LIGO/Virgo/KAGRA O4-like detector network. Unlike previous methods that treat detection as a binary classification problem, this approach frames it as a density estimation problem.

Density Estimation via Bayes' Theorem: The method models the probability density of parameters for detected events, $p(\theta|D)$ , using an NF trained exclusively on the detected events from a public injection dataset. The detection probability is then derived via Bayes' theorem:
$P(D|\theta) = \frac{p(\theta|D)P(D)}{p(\theta)}$
Here, $p(\theta)$ is the known injection population model, and $P(D)$ is the overall detection efficiency (a scalar constant derived from the injection campaign). The NF learns the complex, high-dimensional (13-parameter) distribution of detected events, $p(\theta|D)$ .
Model Architecture: The authors utilize a triangular spline flow, which combines monotonic rational-quadratic splines (acting independently on dimensions) with learned lower-triangular affine maps (introducing dependencies). This architecture allows for efficient computation of the Jacobian determinant, making the density evaluation exact and fast. The model is trained on approximately 500,000 detected events from the O4a and O4b injection campaigns using the JAX/FlowJAX framework.
Evaluation Metrics: The emulator's performance is assessed using two-sample tests to compare distributions:
- Marginal Tests: Kolmogorov-Smirnov (KS) and Anderson-Darling (AD) tests on 1D distributions.
- High-Dimensional Tests: PQMass (using Voronoi tessellation) and a Classifier Two-Sample Test (C2ST) using a fully connected neural network to distinguish between injection data and NF samples.

Key Contributions

Generative Emulator: The paper introduces a generative NF emulator that approximates the selection function by learning the density of detected events rather than a decision boundary. This allows for the generation of synthetic detected events and provides a diagnostic tool via parameter distribution recovery.
Efficiency and Precision: The method offers a computationally efficient alternative to MC-IS. It avoids the "curse of dimensionality" and sample sparsity issues associated with MC-IS when evaluating population models with sharp features.
Public Release: The authors provide a publicly available Python package ("graviflows") containing the training code, scripts for retraining on future campaigns (e.g., O4c), and the fully trained models for O4a and O4b.

Results

Distribution Recovery: The NF successfully reproduces the joint and marginal parameter distributions of the injection test data. Two-sample tests (KS, AD, PQMass, C2ST) yield p-values indicating that the null hypothesis (that NF samples and test data come from the same distribution) cannot be rejected.
Detection Efficiency: The emulator's estimate of detection efficiency for the injection population model matches the MC-IS estimate within $1\sigma$ ( $\approx 1.12 \times 10^{-3}$ ).
Robustness to Population Models: When tested against various population models (including those with varying spin magnitude widths), the NF emulator provides stable estimates of detection efficiency. In contrast, MC-IS estimates exhibit high variance and instability for models with narrow features where $N_{eff}$ drops significantly.
Over-predictions: A small fraction (< 0.25%) of predictions exceed a probability of 1 (over-predictions). The authors find these occur primarily in regions of parameter space where detection is nearly guaranteed (low redshift, high mass) and are attributed to finite training data and model capacity. They are clipped to 1 in the implementation and deemed not to pose a significant problem.

Significance and Claims
The paper claims that this NF-based approach provides a fast, accurate, and flexible tool for evaluating the selection function in GW data analysis. Its primary significance lies in stabilizing hierarchical Bayesian population inference by providing precise detection efficiency estimates without the high variance associated with MC-IS for complex population models. The authors emphasize that the method requires minimal preprocessing, no feature engineering, and achieves training times of roughly 2000 epochs (under one hour on modern GPUs). While not perfect due to minor over-predictions, the emulator effectively interpolates between training data points, offering a robust solution for current and future observing runs (O4 and beyond). The authors explicitly state that the method is intended to replace or augment traditional injection campaigns for the purpose of rapid selection function evaluation in population studies.

More like this