A Neural Network Listens to Black Holes ✛ ✎

Original: "Hierarchical Population Inference with Normalizing Flows for Binary Black Holes"

· Leonardo Iampieri, Simone Mastrogiovanni

arXiv:2609.11885v1 · 2026-09-10 · arXiv non-exclusive · independent analysis · 1 min

Scientists have taught a neural network to reconstruct black hole masses from gravitational waves without assuming the shape of the distribution in advance.

A translator learns a language from conversations, not from a dictionary: they hear fragments of phrases and noise, yet gradually grasp which words matter. That is how a neural network learns from gravitational waves, the tremor of space produced by merging black holes.

Every signal arrives with interference, so we do not know the exact mass of a black hole, only a set of possible values. The neural network takes this uncertainty into account.

Previously, scientists fixed the shape of the mass distribution in advance, the way a tailor cuts a suit to three sizes. If the real shape is different, the model gets it wrong. The new neural network starts with a smooth hill and kneads it until it matches the data. At the same time, it remembers that we cannot see weak signals and applies a correction.

On real data, the neural network found two peaks in the mass distribution: around 10 and 35 solar masses. This agrees with earlier analyses, but now the shape was not imposed from the start. The same method can be used for other tasks where we see only part of the objects, for example counting exoplanets.

Whole graph →

The injection campaign for GWTC-5.0 generated 1.57 billion signals, of which only 1.34 million passed the detection threshold, corresponding to roughly one survivor per thousand generated.

\mathcal{L}(\{d_i\}|\phi) \propto \prod_{i=1}^{C} \frac{\int p(d_i|\theta) q_\phi(\theta) d\theta}{\alpha(\phi)}

The probability of the data given a population density q_φ equals the product over events of the ratio of the integral of the event likelihood times the population density to the selection factor α(φ).

\Delta I_{\tilde{S}} = I_{\mathrm{rec}} - I_{\mathrm{ref}}

The difference in mutual information between the reconstructed density and its reference version without redshift dependence, both restricted to the supported region.

Concepts

normalizing flows Hardy–Weinberg principle Wright–Fisher model bayesian hierarchical model gravitational lensing modeling posterior distribution importance sampling

I am the author

A Neural Network Listens to Black Holes ✛ ✎